Load model selection method and system based on time-space characteristic evolution rule quantitative evaluation, storage medium and electronic equipment

Abstract

The invention discloses a load model selection method and system based on time-space characteristic evolution rule quantitative evaluation, a storage medium and electronic equipment, and the method comprises the steps: obtaining multi-business-state time sequence data for load modeling, and constructing the time sequence characteristics of the multi-business-state time sequence data; performing quantitative preliminary screening on the multi-business-state time sequence data through a constructed dynamic envelope channel, and identifying a parameter drift event; based on the multi-business-state time sequence data, the time sequence features and the parameter drift event, constructing a comprehensive feature vector; obtaining a prediction error of the comprehensive feature vector through a trained LightGBM regression model, and judging a parameter drift event through the prediction error; and when the parameter drift event is judged to be valid, updating a load model in a preset scenarized candidate load model library based on a preset rule, and generating an optimal load model decision sequence.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of large power grid safety and stability and simulation modeling, and more particularly to a load model selection method and system based on spatiotemporal characteristic evolution law quantitative evaluation, a storage medium and an electronic device. BACKGROUND

[0002] With the construction of new power systems, the structure of the load side has undergone fundamental changes: large-scale grid-connected distributed photovoltaic, wind power, and other new loads such as electric vehicle charging piles and smart home have continuously increased their share, making the diversity, randomness, time-varying nature of the load, and the active and power electronic features of the load side increasingly prominent.

[0003] Traditional load modeling generally uses a static aggregation method to represent the load as a combination of constant impedance (Z), constant current (I), constant power (P), and the proportion of induction motors, supplemented by frequency static characteristic factors. The selection of such model parameters is usually based on a small amount of data from a typical day or a typical season, and the update cycle is as long as several months or years. The model is solidified and cannot adapt to the rapid dynamic changes in load characteristics under new power systems, resulting in large errors in power grid simulation calculations and seriously affecting the accuracy of scheduling plans and the effectiveness of stability control strategies.

[0004] In the prior art, to solve the model mismatch problem, some research on dynamically updating models has emerged, but all have significant defects:

[0005] (1) Sensitivity and accuracy contradiction: using statistical process control with a fixed time window or simple threshold method for parameter mutation detection, either insensitive to slow drift or produces a large number of "false drift" false positives for short-term fluctuations.

[0006] (2) Lack of intelligence in model update decision: even if parameter changes are detected, subsequent model selection relies heavily on human experience, which is time-consuming and labor-intensive, and highly subjective, making it difficult to ensure the adaptability of the selected model to future periods.

[0007] (3) Technical isolation and lack of system integration: related technologies have not been effectively integrated. The application of natural wave band algorithms in the financial field, efficient machine learning tools such as LightGBM, and rolling backtesting and Reward mechanisms in quantitative investment have not been seen in dynamic optimization of load models.

[0008] Therefore, a technology is needed to select a load model based on spatiotemporal characteristic evolution law quantitative evaluation. SUMMARY

[0009] The present invention provides a method and system for selecting load models based on quantitative evaluation of spatiotemporal characteristic evolution laws, in order to solve the problem of how to select load models based on quantitative evaluation of spatiotemporal characteristic evolution laws.

[0010] To address the aforementioned problems, this invention provides a load model selection method based on quantitative evaluation of spatiotemporal characteristic evolution laws, the method comprising:

[0011] Acquire multi-business time-series data for load modeling, and construct the time-series features of the multi-business time-series data;

[0012] By constructing a dynamic envelope channel, the multi-business time series data is subjected to natural band-based quantitative analysis to identify parameter drift events;

[0013] Based on the multi-business time-series data, the time-series features, and parameter drift events, a comprehensive feature vector is constructed.

[0014] The prediction error of the comprehensive feature vector is obtained by training the LightGBM regression model, and the parameter drift event is determined by the prediction error.

[0015] When the parameter drift event is determined to be valid, the load model in the preset scenario-based candidate load model library is selected based on preset rules, and the updated load model is taken as the optimal load model for the corresponding scenario.

[0016] Preferably, the multi-business time-series data includes: basic measurement data and load characteristic data;

[0017] The basic measurement data include: total active power (P), total reactive power (Q), average temperature (T), humidity (C), solar radiation intensity (W), and distributed power penetration rate (D) of the entire network or region;

[0018] The load characteristic data includes: motor trend data, static load trend data, load frequency characteristic trend data, load classification composition trend data, detailed electrical load equipment composition trend data, government and industry load classification composition trend data, and distributed power supply trend data.

[0019] The time-series characteristics include: hour, weekday, weekend or holiday.

[0020] Preferably, the method further includes processing the multi-business time-series data, including:

[0021] The isolation forest algorithm was used to remove outliers from the multi-business time-series data.

[0022] Spline interpolation based on spatiotemporal correlation is used to fill in the missing values ​​of the multi-business time series data;

[0023] The time-series data from the various business sectors are standardized.

[0024] Preferably, the step of performing natural band-based quantization analysis on the multi-business time-series data through the constructed dynamic envelope channel to identify parameter drift events includes:

[0025] Backtrack data from multiple preset sampling points within a preset time period to identify the natural peaks and troughs of each sampling point;

[0026] Connecting adjacent identified natural peaks forms an upper rail; connecting adjacent identified troughs forms a lower rail; the upper rail and the lower rail constitute a dynamic envelope channel;

[0027] If the parameter value at the current sampling point is greater than the resistance level of the upper rail of the most recent band, or less than the support level of the lower rail of the most recent band, then a potential parameter drift is determined to have occurred.

[0028] Preferably, the comprehensive feature vector includes: core event features, time-series periodic features, real-time environment features, and system state features.

[0029] Preferably, the step of obtaining the prediction error of the comprehensive feature vector through the trained LightGBM regression model, and determining the parameter drift event through the prediction error, includes:

[0030] The comprehensive feature vector is input into the trained LightGBM regression model to obtain the prediction error of modeling the current load over a future preset time period.

[0031] When the prediction error is 1.5%, the parameter drift event is considered valid.

[0032] Preferably, when the parameter drift event is determined to be valid, the load model in the preset scenario-based candidate load model library is optimized based on preset rules, and the updated load model is taken as the optimal load model for the corresponding scenario, including:

[0033] A pre-built library of scenario-based candidate load models containing multiple typical scenarios;

[0034] Starting from the current moment, each scenario-based candidate load model is substituted into the multi-business time series data within the future backtesting window for quantitative backtesting, and the root mean square error (RMS_ERR_i) of each scenario-based candidate load model within the backtesting window is calculated.

[0035] The Reward_i score for the i-th scenario-based candidate load model is calculated based on the backtest root mean square error RMS_ERR_i:

[0036] Reward_i = 1 / (1+RMS_ERR_i);

[0037] The scenario-based candidate load model with the highest Reward_i score is selected as the optimal load model.

[0038] Preferably, the method further includes:

[0039] Record the optimal load model for each time period to generate a sequence of optimal load models for the whole year (96×365).

[0040] Generate a visualization view of the optimal load model sequence;

[0041] Generate an evolution analysis report of the load model according to a preset cycle.

[0042] Preferably, the method further includes:

[0043] Based on new multi-business time-series data, the scenario-based candidate load model library is evaluated periodically, and scenario-based candidate load models with continuously declining performance are marked or deleted.

[0044] New scenario-based candidate load models are generated through cluster analysis.

[0045] In another aspect, the present invention provides a load model selection system based on quantitative evaluation of spatiotemporal characteristic evolution laws, the system comprising:

[0046] An initial unit is used to acquire multi-business time-series data for load modeling and to construct the time-series features of the multi-business time-series data;

[0047] The identification unit is used to perform natural band-based quantitative analysis on the multi-business time series data through the constructed dynamic envelope channel to identify parameter drift events;

[0048] The construction unit is used to construct a comprehensive feature vector based on the multi-business time-series data, the time-series features, and parameter drift events;

[0049] The judgment unit is used to obtain the prediction error of the comprehensive feature vector through the trained LightGBM regression model, and to judge the parameter drift event through the prediction error.

[0050] The result unit is used to select the optimal load model from the preset scenario-based candidate load model library based on preset rules when the parameter drift event is determined to be valid, and to use the updated load model as the optimal load model for the corresponding scenario.

[0051] According to another aspect of the present invention, the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the program is executed by a processor, it implements steps such as a load model selection method based on a quantitative evaluation of spatiotemporal characteristic evolution laws.

[0052] According to another aspect of the present invention, the present invention provides an electronic device, comprising:

[0053] The aforementioned computer-readable storage medium; and

[0054] One or more processors for executing a program in the computer-readable storage medium.

[0055] This invention provides a load model selection method based on the quantitative assessment of spatiotemporal characteristic evolution laws. The method includes: acquiring multi-sector time-series data for load modeling and constructing the time-series features of the multi-sector time-series data; performing natural band-based quantitative analysis on the multi-sector time-series data through a constructed dynamic envelope channel to identify parameter drift events; constructing a comprehensive feature vector based on the multi-sector time-series data, time-series features, and parameter drift events; obtaining the prediction error of the comprehensive feature vector through a trained LightGBM regression model, and determining parameter drift events based on the prediction error; when a parameter drift event is determined to be valid, selecting a load model from a pre-set scenario-based candidate load model library based on preset rules, and using the updated load model as the optimal load model for the corresponding scenario. This invention provides a method and system for the quantitative assessment and adaptive optimization of load model evolution laws based on massive time-series data and machine learning algorithms, achieving key parameter drift identification and automatic decision-making for the optimal model sequence. Attached Figure Description

[0056] Exemplary embodiments of the present invention can be more fully understood by referring to the following figures:

[0057] Figure 1 This is a flowchart of a load model selection method based on a quantitative assessment of spatiotemporal characteristic evolution law according to a preferred embodiment of the present invention.

[0058] Figure 2 This is a system overall architecture block diagram according to a preferred embodiment of the present invention;

[0059] Figure 3 This is a flowchart of the key algorithm according to a preferred embodiment of the present invention;

[0060] Figure 4 This is a schematic diagram of a 15-minute load-temperature K-line according to a preferred embodiment of the present invention;

[0061] Figure 5This is a 4:30 point natural band drift marker diagram according to a preferred embodiment of the present invention;

[0062] Figure 6 This is a schematic diagram of the resistance and support levels of the load characteristic time series data according to a preferred embodiment of the present invention;

[0063] Figure 7 This is a schematic diagram of the peak and upward drift event marking according to a preferred embodiment of the present invention;

[0064] Figure 8 This is a schematic diagram of the trough and downward drift event marking according to a preferred embodiment of the present invention;

[0065] Figure 9 This is a scatter plot of the LightGBM event-error response surface according to a preferred embodiment of the present invention;

[0066] Figure 10 This is a schematic diagram of a "morning rush hour" drift event marker for an "event scenario" according to a preferred embodiment of the present invention;

[0067] Figure 11 This is a schematic diagram of drift event marking for a "summer air conditioning season" in a "climate scenario" according to a preferred embodiment of the present invention;

[0068] Figure 12 This is a schematic diagram of a drift event marker for a "high photovoltaic treatment" in a "special scenario" according to a preferred embodiment of the present invention;

[0069] Figure 13 This is a reward score curve for the 8-week model according to a preferred embodiment of the present invention;

[0070] Figure 14 This is a schematic diagram of the Reward score for a "summer air conditioning season" model based on a preferred embodiment of the present invention, targeting "weekdays".

[0071] Figure 15 This is a schematic diagram of the "Reward" model for "weekends" according to a preferred embodiment of the present invention;

[0072] Figure 16 This is a preferred embodiment of the 96×365 optimal model color block calendar diagram for the whole year according to the present invention;

[0073] Figure 17 This is a schematic diagram of a color block calendar for a June "summer air conditioning season model" according to a preferred embodiment of the present invention;

[0074] Figure 18 This is a schematic diagram of a color block calendar representing a "special holiday model" for the whole year, according to a preferred embodiment of the present invention.

[0075] Figure 19 This is a schematic diagram of a designated single load model cross-section Tooltip according to a preferred embodiment of the present invention;

[0076] Figure 20 The system structure diagram is selected for the load model based on the quantitative evaluation of spatiotemporal characteristic evolution law according to the preferred embodiment of the present invention. Detailed Implementation

[0077] Exemplary embodiments of the invention will now be described with reference to the accompanying drawings. However, the invention may be embodied in many different forms and is not limited to the embodiments described herein. These embodiments are provided to fully and completely disclose the invention and to fully convey its scope to those skilled in the art. The terminology used in the exemplary embodiments illustrated in the drawings is not intended to limit the invention. In the drawings, the same units / elements are referred to by the same reference numerals.

[0078] Unless otherwise stated, the terms used herein (including technical terms) have their common meaning as understood by one of ordinary skill in the art. Furthermore, it is understood that terms defined in commonly used dictionaries should be understood to have a meaning consistent with the context of their relevant field, and not to be interpreted as having an idealized or overly formal meaning.

[0079] Figure 1 This is a flowchart of a load model selection method based on a quantitative assessment of spatiotemporal characteristic evolution laws according to a preferred embodiment of the present invention.

[0080] This invention provides an adaptive optimization method and system for load models based on quantitative evaluation of evolution laws, aiming to achieve:

[0081] (1) Capture substantial abrupt changes in load model parameters in near real time (lag <30 minutes);

[0082] (2) Using machine learning techniques, the rate of invalid "false drift" alerts can be reduced by more than 70%;

[0083] (3) Automatically output the optimal load model sequence of 96 points / day × 365 days throughout the year, reducing the time for manual intervention from several weeks to minutes;

[0084] (4) The RMS error of the dynamic simulation of the power system was stabilized and reduced from 2.8% at the baseline level to below 1.2%;

[0085] (5) Reduce scheduling conservatism and rotating reserve caused by model inaccuracy, and achieve annual scheduling cost savings of ≥24 million yuan.

[0086] This invention proposes a quantitative evaluation and adaptive optimization technique for load model evolution based on massive time-series data and machine learning algorithms. Its core lies in constructing an automated closed-loop method of "detection-filtering-backtesting-decision-feedback," and the software system upon which this method relies includes the following logical modules in sequence:

[0087] (1) Data preprocessing and feature engineering module: responsible for accessing, cleaning, aligning and extracting features from multi-source heterogeneous data, especially using the isolated forest algorithm for outlier detection and filling in missing data based on spatiotemporal correlation;

[0088] (2) Parameter drift screening module: Based on the improved natural band algorithm, quantitative analysis and initial screening of change points are performed on the time series trend data of key parameters of the critical load model.

[0089] (3) False drift intelligent filtering module: Using the LightGBM machine learning model, a "event-error" mapping relationship between high-dimensional features and model errors is constructed to evaluate the credibility of the initial screening alarm and filter it, effectively distinguishing noise fluctuations from structural drift;

[0090] (4) Multi-model rolling backtesting and optimization module: Maintain a dynamically expanding "scenario-based candidate model library" and, based on the rolling time window and a custom reward function, perform a forward-looking evaluation and optimization of the adaptability of each candidate model within a future time window;

[0091] (5) Model Sequence Management and Visual Decision Module: Records and outputs the optimal model decision sequence, and provides intuitive decision support to operators through a color block calendar and interactive Tooltip visualization interface;

[0092] (6) Feedback learning and model library update module: Based on new data and adaptive feedback accumulated during long-term operation, the candidate model library is evaluated and optimized regularly to achieve the self-evolution and continuous optimization of the system.

[0093] like Figure 1 As shown, this invention provides a load model selection method based on quantitative evaluation of spatiotemporal characteristic evolution laws, characterized in that the method includes:

[0094] Step 101: Obtain multi-business time-series data for load modeling and construct the time-series features of the multi-business time-series data;

[0095] Preferably, the multi-business time series data includes: basic measurement data and load characteristic data;

[0096] Basic measurement data include: total active power (P), total reactive power (Q), average temperature (T), humidity (C), solar radiation intensity (W), and distributed power penetration rate (D) of the entire network or region;

[0097] Load characteristic data includes: motor trend data, static load trend data, load frequency characteristic trend data, load classification composition trend data, detailed electrical load equipment composition trend data, government and industry load classification composition trend data, and distributed power supply trend data.

[0098] The time sequence characteristics include: hour, weekday, weekend or holiday.

[0099] Preferably, the method further includes processing multi-business time-series data, including:

[0100] The isolation forest algorithm was used to remove outliers from time-series data of multiple business formats.

[0101] Spline interpolation based on spatiotemporal correlation is used to fill in missing values ​​in multi-business time-series data;

[0102] Standardization processing is performed on time-series data from multiple business sectors.

[0103] In step 101, the present invention performs data preprocessing and feature engineering:

[0104] This invention acquires 15-minute granularity load modeling-related multi-sector time-series data from control cloud platforms, data middleware, climate and meteorological systems, and distributed generation prediction systems, and performs outlier removal, missing value imputation, and standardization on the raw data. Simultaneously, it constructs time-series features, such as hourly and weekday / weekend / holiday markers. The multi-sector time-series data includes at least:

[0105] Basic measurement data: total active power (P), total reactive power (Q), average temperature (T), humidity (C), solar radiation intensity (W), and distributed power penetration rate (D) of the entire network or region.

[0106] Load characteristic data: Load characteristic data generated through automatic load modeling, such as: motor trend data, static load trend data (constant impedance Z, constant current I, constant power P), load frequency characteristic trend data, load classification and composition trend data (industrial, agricultural, commercial, residential), detailed electrical load equipment composition trend data (large industrial motors, small industrial motors, variable frequency air conditioners, charging piles, etc.), government industry load classification and composition trend data (primary industry, secondary industry, tertiary industry, ordinary urban and rural residents), distributed power source trend data (Class A new energy, Class B new energy, doubly fed wind turbines), etc.

[0107] The isolated forest algorithm was used to detect and remove outliers. Missing data were filled using spline interpolation based on spatiotemporal correlation and then standardized.

[0108] Step 102: Perform natural band-based quantitative analysis on multi-business time series data through the constructed dynamic envelope channel to identify parameter drift events;

[0109] Preferably, parameter drift events are identified by performing natural band-based quantitative analysis on multi-business time-series data through a constructed dynamic envelope channel, including:

[0110] Backtrack data from multiple preset sampling points within a preset time period to identify the natural peaks and troughs of each sampling point;

[0111] Connecting adjacent identified natural peaks forms the upper rail; connecting adjacent identified troughs forms the lower rail; the upper and lower rails together form a dynamic envelope channel.

[0112] If the parameter value at the current sampling point is greater than the resistance level of the upper rail of the most recent band, or less than the support level of the lower rail of the most recent band, then a potential parameter drift is determined to have occurred.

[0113] In step 102 of this invention, parameter drift is initially screened based on the natural band quantization criterion:

[0114] This invention independently performs natural band analysis on the time series of key load model characteristic parameters (such as ZIP coefficient, motor ratio M, frequency factor F, load composition trend, etc.), specifically:

[0115] (1) Backtrack the data of the past 30 sampling points (i.e. 7.5 hours) to identify all the natural peaks (local highs) and troughs (local lows) in between. Define: if the value of the middle point in three consecutive sampling points is greater than (or less than) its two adjacent points, then the point is a potential peak (or trough).

[0116] (2) In this way, all natural peaks and troughs in the past 30 points (7.5 hours) are identified, adjacent peaks are connected to form an "upper rail", and adjacent troughs are connected to form a "lower rail", thus forming a dynamic envelope channel.

[0117] (3) Drift event triggering criterion: If the current parameter value F(t) breaks through the upper resistance level HHV(F,30)[t-2] of the most recent wave band upwards, or breaks through the lower support level LLV(F,30)[t-2] of the most recent wave band downwards, then a potential parameter drift is determined to have occurred, and an initial drift event flag Event_flag = 1 is generated. That is, the criterion is: F(t) > HHV(F,30)[t-2] or F(t) > HHV(F,30)[t-2]. <LLV(F,30)[t-2]。

[0118] Step 103: Construct a comprehensive feature vector based on multi-business time series data, time series features, and parameter drift events;

[0119] Preferably, the comprehensive feature vector includes: core event features, time-series periodic features, real-time environment features, and system state features.

[0120] Step 104: Obtain the prediction error of the comprehensive feature vector through the trained LightGBM regression model, and use the prediction error to determine the parameter drift event;

[0121] Preferably, the prediction error of the comprehensive feature vector is obtained through the trained LightGBM regression model, and the parameter drift event is determined based on the prediction error, including:

[0122] The comprehensive feature vector is input into the trained LightGBM regression model to obtain the prediction error of using the current load to model the future within a preset time period.

[0123] When the prediction error is 1.5%, the parameter drift event is considered valid.

[0124] In step 103 of this invention, pseudo-drift filtering is based on the LightGBM response surface:

[0125] This invention aims to distinguish between "noise fluctuations" and "structural drift" of parameters, and to construct a lightweight LightGBM regression model, as described in detail in the core algorithm formula (2), wherein:

[0126] The comprehensive feature vector Φ(t) includes the initial drift event flag (Event_flag), current temperature (T), holiday flag (Holiday), and real-time distributed permeability (D), and extends this by incorporating the historical backtesting root mean square error (RMS) trend. The RMS is used for model backtesting and optimization, objectively evaluating the adaptability of each candidate model in the "near future," and is a direct basis for calculating the reward score, answering the question, "Which model will perform best in the upcoming period?"

[0127] ① Core event characteristics: Event_flag, Event_magnitude

[0128] ② Time-series periodic characteristics: Holiday, Hour of day, Day of week, Season

[0129] ③ Real-time environment characteristics: T, T_trend, Humidity, Solar_irradiance, D

[0130] ④ System state characteristics: RMS, Current_Model_Age

[0131] Prediction objective: The model was trained to generate the prediction error ErrPred for "continuing to use the old model's adaptability over the next 2 hours (8 sampling points)".

[0132] ErrPred = LGBM_Model(Φ(t))

[0133] It is predicted by the LightGBM model and uses the prediction error of the current model in future periods for false drift intelligent filtering. It does not need to actually perform a time-consuming simulation, but quickly predicts the risk of "maintaining the status quo" and answers "Is it necessary to start the time-consuming model backtesting and update process?" ErrPred acts as an efficient gatekeeper, ensuring that only truly important events trigger the subsequent complex optimization process, thereby achieving efficient system operation while ensuring accuracy.

[0134] The logical closed loop of both RMS and ErrPred:

[0135] ① Offline training: The LightGBM model is able to predict ErrPred because it has learned from massive amounts of historical data and learned how to associate the feature pattern Φ(t) with the RMS_ERR that will eventually occur.

[0136] ② Online Application: When the system detects a parameter drift (event), it first performs a rapid risk assessment using low-cost ErrPred. Only when the risk is sufficiently high (ErrPred > 1.5%) is it worthwhile to initiate the high-cost model backtesting and optimization process based on the real RMS_ERR.

[0137] (3) Decision-making mechanism: Input the real-time features into the trained LightGBM model to obtain the prediction error ErrPred. Set a dynamic decision-making mechanism: If ErrPred>1.5%, the drift is confirmed as a valid drift event and the model update process is started; otherwise, it is judged as a false drift, the system maintains the original model, and ignores the alarm.

[0138] Step 105: When the parameter drift event is determined to be valid, the load model in the preset scenario-based candidate load model library is selected based on preset rules, and the updated load model is taken as the optimal load model for the corresponding scenario.

[0139] Preferably, when the parameter drift event is determined to be valid, a load model from a pre-set scenario-based candidate load model library is selected based on preset rules, and the selected load model is taken as the optimal load model, including:

[0140] A pre-built library of scenario-based candidate load models containing multiple typical scenarios;

[0141] Starting from the current moment, each scenario-based candidate load model is substituted into the multi-business time series data within the future backtesting window for quantitative backtesting, and the root mean square error (RMS_ERR_i) of each scenario-based candidate load model within the backtesting window is calculated.

[0142] The Reward_i score for the i-th scenario-based candidate load model is calculated based on the backtesting root mean square error RMS_ERR_i:

[0143] Reward_i = 1 / (1+RMS_ERR_i);

[0144] The scenario-based candidate load model with the highest Reward_i score is selected as the optimal load model.

[0145] In step 104 of this invention, candidate model rolling simulation backtesting and reward selection are performed:

[0146] Once a valid drift event is confirmed, the system initiates a model optimization process:

[0147] (1) Candidate model library: A library containing multiple typical scenario load models is pre-set. These models are obtained based on historical big data clustering analysis and cover: time scenarios (morning peak, evening peak, night trough, noon trough), climate scenarios (summer high temperature season, winter heating season, spring and autumn average), special scenarios (special holidays, large photovoltaic power generation, small photovoltaic power generation), etc., and can be dynamically expanded according to actual application scenarios.

[0148] (2) Rolling backtesting: Starting from the current time, each candidate model is substituted into the time series data of the next 4 hours (16 sampling points) for rapid quantization backtesting;

[0149] (3) Reward score calculation: For each candidate model i, calculate its backtest root mean square error RMS_ERR_i within the entire backtesting window, and calculate its reward value accordingly: Reward_i=1 / (1+RMS_ERR_i);

[0150] (4) Optimal Model Selection: The candidate model with the highest Reward_i value is selected as the official load model for system simulation and decision analysis in the next time period. Optimal Decision:

[0151] optimal_model=argmax(Reward_i)

[0152] Specifically, if the Reward score of all models is below 0.5 (i.e., the usability threshold), a model library warning is triggered, indicating that a new model needs to be introduced.

[0153] Preferably, the method further includes:

[0154] Record the optimal load model for each time period to generate a sequence of optimal load models for the whole year (96×365).

[0155] Generate a visual view of the optimal load model sequence;

[0156] Generate an evolution analysis report of the load model according to a preset cycle.

[0157] In step 105 of this invention, optimal sequence generation and visualization are performed:

[0158] This invention records the optimal model selected for each time period, forming a 96×365 model sequence for the entire year, and outputs it in a structured format for direct use by commercial simulation software such as PSD-BPA and PSASP. Simultaneously, it generates a color-block calendar visualization: the horizontal axis represents the 96 time periods of a day, the vertical axis represents the dates of the year, and each color block represents the optimal model type used in that time period, with different colors representing different models. It supports mouse hover (tooltip) interaction, instantly displaying the specific model parameters, reward score, and key environmental variables for that time period. Furthermore, the system automatically generates a weekly model evolution analysis report to help operators understand macro-level patterns.

[0159] Preferably, the method further includes:

[0160] Based on new multi-business time-series data, the scenario-based candidate load model library is evaluated regularly, and scenario-based candidate load models with continuously declining performance are marked or deleted.

[0161] New scenario-based candidate load models are generated through cluster analysis.

[0162] In step S106, this invention performs feedback learning and dynamic updating of the model library:

[0163] This invention automatically performs feedback learning once a month: using new data accumulated over the past month, it re-evaluates the generalization performance of each candidate model. Models with continuously declining performance are marked or eliminated; simultaneously, new load patterns are discovered through cluster analysis, and corresponding new models are added to the candidate library, achieving self-evolution and dynamic optimization of the model library.

[0164] The core algorithm formula of this invention includes:

[0165] (1) Waveband Breakout Judgment Formula:

[0166] drift_event=(F(t)>HHV(F,30)[t-2])OR(F(t) <LLV(F,30)[t-2])

[0167] LightGBM prediction model:

[0168] ErrPred = LGBM_Model(Φ(t))

[0169] Φ(t)=[Event_flag,Event_magnitude,Holiday,Hour_of_day,Day_of_week,Season,T,T_trend,Humidity,Solar_irradiance,D,Recent_RMS,Current_Model_Age]

[0170] Where Φ(t) represents the comprehensive feature vector constructed at time t, and ErrPred is the prediction error of the model's "adaptability to continue using the current load model in the next 2 hours (8 sampling points)".

[0171] Step 1: Engineering construction of eigenvector Φ(t)

[0172] The eigenvector Φ(t) is a high-dimensional vector designed to comprehensively capture all factors affecting the accuracy of the load model. It consists of the following feature set:

[0173] ①Key characteristics of the event:

[0174] Event_flag: Initial drift event flag generated by step S102 (1 indicates potential drift, 0 indicates none).

[0175] Event_magnitude: Event intensity, defined as |F(t) - most recent orbital value| / orbital width, quantifying the significance of a breakout.

[0176] ②Time series and periodic characteristics:

[0177] Holiday: Holiday markers (e.g., regular workday = 0, weekend = 1, statutory holiday = 2).

[0178] Hour_of_day, Day_of_week: Hours of the day (0-23) and days of the week (0-6), encoded with sine and cosine to capture periodicity.

[0179] Season: Seasonal markers (1-4).

[0180] ③ Real-time environmental characteristics:

[0181] T: Current temperature.

[0182] T_trend: Temperature trend, such as the average slope over the past 2 hours.

[0183] Humidity, Solar irradiance: Humidity and solar irradiance.

[0184] D: Real-time distributed power penetration rate.

[0185] ④ System status and historical performance characteristics:

[0186] RMS: The moving average RMS error of the current load model over the past 4 hours, reflecting the recent health of the model.

[0187] Current_Model_Age: The duration for which the current model has been used continuously.

[0188] Step 2: Training and building the LGBM_Model

[0189] ① Model selection: The regression model under the LightGBM framework is adopted because it can efficiently handle the above mixed types of tabular data and automatically capture the complex interaction effects between features.

[0190] ② Training data preparation:

[0191] Source: Historical data spanning 1-2 years.

[0192] Sample construction: For each historical time t, a feature vector Φ_historical(t) can be constructed.

[0193] Label: The label Y of this sample is the actual RMS error that would occur within the next 2 hours at historical time t, if the model at that time were used. This is a post-hoc calculated true value.

[0194] ③Key training parameters:

[0195] Objective function: For regression tasks, regression_l1 (mean absolute error) or regression (mean squared error) are usually selected.

[0196] Important hyperparameters:

[0197] num_leaves: 31-127 (controls model complexity)

[0198] learning_rate: 0.05-0.1 (controls the learning speed)

[0199] feature_fraction: 0.8 (The proportion of features used in each iteration, enhancing robustness)

[0200] lambda_l1, lambda_l2: L1 and L2 regularization terms to prevent overfitting.

[0201] Step 3: Online Decision-Making Mechanism

[0202] After the model is trained, it is integrated into the online system, and its decision-making process is as follows:

[0203] ① Real-time inference: At each sampling point t_now, the system constructs the current feature vector Φ(t_now) in real time and inputs it into the trained LGBM_Model.

[0204] ② Error prediction: The model outputs the prediction error ErrPred.

[0205] ③ Intelligent decision-making:

[0206] If ErrPred > θ (where the threshold θ = 1.5%):

[0207] THEN: Determined as valid drift. The system determines that continuing to use the current model will result in unacceptable accuracy loss and immediately triggers the subsequent model simulation backtesting and optimization process (step S104).

[0208] ELSE:

[0209] THEN: Determined as a false drift or acceptable fluctuation. The system maintains the current load model and stores {Φ(t_now), the actual future error} from this inference as a new data point in the database for future incremental training of the model.

[0210] Reward scoring function:

[0211] Reward_i = 1 / (1 + RMS_ERR_i)

[0212] (4) Model optimization decision:

[0213] optimal_model=argmax(Reward_i)foriin candidate_models

[0214] The present invention will be further described in conjunction with a preferred embodiment deployed in a provincial power grid.

[0215] Implementation environment: Automatic load modeling platform of a provincial dispatch center of State Grid Corporation of China

[0216] Hardware configuration: 4 cloud virtual servers, each with a 96-core CPU and 256GB DDR4 memory.

[0217] Software environment: NingSi Security Operating System V6.0.80, Qt 5.8.0, and DM8 database.

[0218] Implementation steps:

[0219] 1) Data Access and Preparation: Access the province's grid with historical 15-minute grid model data, grid operation data, climate and meteorological data, and distributed new energy ledgers and operation data for three years from 2022 to 2024 through the control cloud platform. High-frequency data is synchronized once a day, and low-frequency data is synchronized once a week to ensure that the data requirements for active distribution network topology analysis and online integrated statistical method automatic load modeling are met.

[0220] 2) Model training and initialization:

[0221] (1) Using the massive load model parameters accumulated from 2022 to 2024, multiple typical power load model candidate scenarios are determined through cluster analysis, and corresponding candidate load models are generated by aggregation to build a candidate model library;

[0222] (2) Using 2023 data and corresponding simulation error records, a LightGBM model for pseudo-drift filtering was trained.

[0223] 3) Online operation and verification:

[0224] (1) Using the data for the entire year of 2024 as input, run the algorithm of this invention for quantitative backtesting;

[0225] (2) The system completes the calculation for the whole year within 20 minutes, triggering an initial drift alarm 147 times;

[0226] (3) After filtering by the LightGBM model, 41 valid drift events were confirmed, with a false alarm filtering rate of 72.1%;

[0227] (4) The system automatically generated the optimal model sequence of 35,040 points (96×365) throughout the year.

[0228] 4) Effect verification:

[0229] Scenario 1: Summer Major Mode Fault Simulation Analysis

[0230] The generated model sequence was exported as structured data for selection by the PSD-BPA software for large-scale simulation in the summer of 2024. The simulation results were compared with the actual waveform data recorded in the summer of 2024. Results: The simulation RMS error was 1.1%, significantly lower than 2.8% when using the traditional fixed model, fully meeting the accuracy requirement of less than 5% error in the power system safety and stability guidelines.

[0231] Scenario 2: Sudden Changes in Load Characteristics During the National Day Holiday

[0232] During the 2024 National Day holiday, the system detected a trend of increasing proportion of "commercial" and "residential" load components during the evening peak on the day before the holiday, triggering parameter drift. After being confirmed as a valid event by LightGBM filtering, the system switched the model to the "special holiday" model in advance through backtesting, ensuring that the simulation accuracy remained within 1.3% throughout the National Day holiday.

[0233] Scenario 3: High Solar Output Hours in Summer Afternoons

[0234] During a summer midday in a certain provincial power grid, a surge in photovoltaic (PV) output led to rapid changes in load characteristics. Traditional methods, due to update lag, resulted in an error of up to 3.5%. This invention detects a band break in the distributed PV proportion parameter within 15 minutes after PV penetration exceeds 25%. Through intelligent filtering and backtesting optimization, it quickly switches to the "PV power surge" model, stabilizing the simulation error for that period at 1.0%.

[0235] 5) Economic evaluation: Based on the optimization of reserve capacity brought about by the reduction of error, the finance department calculated that the application of this system will save the provincial power grid approximately 24 million yuan in dispatching and operation costs annually.

[0236] The present invention has the following significant advantages:

[0237] (1) This invention is the first to apply natural band analysis technology to load parameter drift detection, which solves the problems of lag and false alarm in the fixed window method and realizes the keen capture of the "trend reversal" of parameters.

[0238] (2) This invention constructs an intelligent filtering mechanism for the "event-error" response surface, and uses the LightGBM model to quantify the complex nonlinear relationship between external operating conditions and model mismatch, fundamentally reducing the false alarm rate (the LightGBM intelligent filtering mechanism reduces the false alarm rate from 72% to 28%).

[0239] (3) The present invention introduces a model prospective optimization strategy based on rolling backtesting and the Reward function, which changes the model selection from "post-fitting" to "pre-prediction", ensuring the optimal adaptability of the selected model in the future period.

[0240] (4) This invention forms a complete automated closed-loop technology chain. From data input to decision output, the entire process does not require manual intervention, which improves modeling efficiency by several orders of magnitude.

[0241] (5) The present invention has significant economic benefits: by improving simulation accuracy, it can significantly reduce the excessive spinning reserve and conservative dispatch schemes set up to cope with uncertainties. According to calculations on a provincial power grid with an annual electricity consumption of 50 billion kWh, the present invention can save approximately RMB 24 million in dispatch costs annually.

[0242] (6) This invention introduces a feedback learning and dynamic update mechanism for the model library, enabling the system to adapt to the long-term evolution of the active distribution network and possess the ability to continuously optimize, fundamentally solving the problem of performance degradation of traditional systems over time.

[0243] (7) The present invention improves the engineering practicality and robustness of the system, and ensures long-term stable operation in real complex power grid environment through detailed exception handling logic and dynamic threshold mechanism.

[0244] Figure 20 The system structure diagram is selected for the load model based on the quantitative evaluation of spatiotemporal characteristic evolution law according to the preferred embodiment of the present invention.

[0245] like Figure 20 As shown, this invention provides a load model selection system based on quantitative evaluation of spatiotemporal characteristic evolution laws, characterized in that the system includes:

[0246] Initial unit 201 is used to acquire multi-business time-series data for load modeling and to construct the time-series features of the multi-business time-series data;

[0247] Preferably, the multi-business time series data includes: basic measurement data and load characteristic data;

[0248] Basic measurement data include: total active power (P), total reactive power (Q), average temperature (T), humidity (C), solar radiation intensity (W), and distributed power penetration rate (D) of the entire network or region;

[0249] Load characteristic data includes: motor trend data, static load trend data, load frequency characteristic trend data, load classification composition trend data, detailed electrical load equipment composition trend data, government and industry load classification composition trend data, and distributed power supply trend data.

[0250] The time sequence characteristics include: hour, weekday, weekend or holiday.

[0251] Preferably, the method further includes processing multi-business time-series data, including:

[0252] The isolation forest algorithm was used to remove outliers from time-series data of multiple business formats.

[0253] Spline interpolation based on spatiotemporal correlation is used to fill in missing values ​​in multi-business time-series data;

[0254] Standardization processing is performed on time-series data from multiple business sectors.

[0255] The identification unit 202 is used to perform natural band-based quantitative analysis on multi-business time series data through the constructed dynamic envelope channel to identify parameter drift events;

[0256] Preferably, the identification unit 202 is used to perform natural band-based quantitative analysis on multi-business time-series data through a constructed dynamic envelope channel to identify parameter drift events, including:

[0257] Backtrack data from multiple preset sampling points within a preset time period to identify the natural peaks and troughs of each sampling point;

[0258] Connecting adjacent identified natural peaks forms the upper rail; connecting adjacent identified troughs forms the lower rail; the upper and lower rails together form a dynamic envelope channel.

[0259] If the parameter value at the current sampling point is greater than the resistance level of the upper rail of the most recent band, or less than the support level of the lower rail of the most recent band, then a potential parameter drift is determined to have occurred.

[0260] Construction unit 203 is used to construct a comprehensive feature vector based on multi-business time series data, time series features, and parameter drift events;

[0261] Preferably, the comprehensive feature vector includes: core event features, time-series periodic features, real-time environment features, and system state features.

[0262] The decision unit 204 is used to obtain the prediction error of the comprehensive feature vector through the trained LightGBM regression model, and to determine the parameter drift event based on the prediction error.

[0263] Preferably, the determination unit 204 is used to obtain the prediction error of the comprehensive feature vector through the trained LightGBM regression model, and to determine the parameter drift event based on the prediction error, including:

[0264] The comprehensive feature vector is input into the trained LightGBM regression model to obtain the prediction error of using the current load to model the future within a preset time period.

[0265] When the prediction error is 1.5%, the parameter drift event is considered valid.

[0266] Result unit 205 is used to select a load model from a preset scenario-based candidate load model library based on preset rules when the parameter drift event is determined to be valid, and to use the updated load model as the optimal load model for the corresponding scenario.

[0267] Preferably, the result unit 205 is used to update the load model in the preset scenario-based candidate load model library based on preset rules when the parameter drift event is determined to be valid, and to use the updated load model as the optimal load model for the corresponding scenario, including:

[0268] A pre-built library of scenario-based candidate load models containing multiple typical scenarios;

[0269] Starting from the current moment, each scenario-based candidate load model is substituted into the multi-business time series data within the future backtesting window for quantitative backtesting, and the root mean square error (RMS_ERR_i) of each scenario-based candidate load model within the backtesting window is calculated.

[0270] The Reward_i score for the i-th scenario-based candidate load model is calculated based on the backtesting root mean square error RMS_ERR_i:

[0271] Reward_i = 1 / (1+RMS_ERR_i);

[0272] The scenario-based candidate load model with the highest Reward_i score is selected as the optimal load model.

[0273] Preferably, the result unit 205 is further configured to:

[0274] Record the optimal load model for each time period to generate a sequence of optimal load models for the whole year (96×365).

[0275] Generate a visual view of the optimal load model sequence;

[0276] Generate an evolution analysis report of the load model according to a preset cycle.

[0277] Preferably, the result unit 205 is further configured to:

[0278] Based on new multi-business time-series data, the scenario-based candidate load model library is evaluated regularly, and scenario-based candidate load models with continuously declining performance are marked or deleted.

[0279] New scenario-based candidate load models are generated through cluster analysis.

[0280] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements steps such as a load model selection method based on a quantitative assessment of spatiotemporal characteristic evolution laws.

[0281] This invention provides an electronic device, comprising:

[0282] The aforementioned computer-readable storage medium; and

[0283] One or more processors for executing programs in a computer-readable storage medium.

[0284] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0285] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0286] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0287] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0288] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0289] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0290] The invention has been described with reference to a few embodiments. However, as will be known to those skilled in the art, and as defined in the appended claims, other embodiments besides those disclosed above fall equivalently within the scope of the invention.

[0291] Generally, all terms used in the claims are to be interpreted according to their ordinary meaning in the art, unless otherwise expressly defined herein. All references to “a / the / the [device, component, etc.]” ​​are openly interpreted as at least one instance of said device, component, etc., unless otherwise expressly stated. The steps of any method disclosed herein need not be performed in the exact order disclosed unless explicitly stated otherwise.

Claims

1. A method for selecting a load model based on quantitative assessment of spatiotemporal characteristic evolution, characterized in that, The method includes: Acquire multi-business time-series data for load modeling, and construct the time-series features of the multi-business time-series data; By constructing a dynamic envelope channel, the multi-business time series data is subjected to natural band-based quantitative analysis to identify parameter drift events; Based on the multi-business time-series data, the time-series features, and parameter drift events, a comprehensive feature vector is constructed. The prediction error of the comprehensive feature vector is obtained by training the LightGBM regression model, and the parameter drift event is determined by the prediction error. When the parameter drift event is determined to be valid, the load models in the preset scenario-based candidate load model library are updated based on preset rules, and the updated load model is selected as the optimal load model for the corresponding scenario.

2. The method according to claim 1, characterized in that, The multi-business time-series data includes: basic measurement data and load characteristic data; The basic measurement data include: total active power (P), total reactive power (Q), average temperature (T), humidity (C), solar radiation intensity (W), and distributed power penetration rate (D) of the entire network or region; The load characteristic data includes: motor trend data, static load trend data, load frequency characteristic trend data, load classification composition trend data, detailed electrical load equipment composition trend data, government and industry load classification composition trend data, and distributed power supply trend data. The time-series characteristics include: hour, weekday, weekend or holiday.

3. The method according to claim 2, characterized in that, The method further includes processing the multi-business time-series data, including: The isolation forest algorithm was used to remove outliers from the multi-business time-series data. Spline interpolation based on spatiotemporal correlation is used to fill in the missing values ​​of the multi-business time series data; The time-series data from the various business sectors are standardized.

4. The method according to claim 1, characterized in that, The step of performing natural band-based quantization analysis on the multi-business time-series data through the constructed dynamic envelope channel to identify parameter drift events includes: Backtrack data from multiple preset sampling points within a preset time period to identify the natural peaks and troughs of each sampling point; Connecting adjacent identified natural peaks forms an upper rail; connecting adjacent identified troughs forms a lower rail; the upper rail and the lower rail constitute a dynamic envelope channel; If the parameter value at the current sampling point is greater than the resistance level of the upper rail of the most recent band, or less than the support level of the lower rail of the most recent band, then a potential parameter drift is determined to have occurred.

5. The method according to claim 1, characterized in that, The comprehensive feature vector includes: core event features, time-series periodic features, real-time environment features, and system state features.

6. The method according to claim 5, characterized in that, The step of obtaining the prediction error of the comprehensive feature vector through the trained LightGBM regression model, and determining the parameter drift event based on the prediction error, includes: The comprehensive feature vector is input into the trained LightGBM regression model to obtain the prediction error of modeling the current load over a future preset time period. When the prediction error is 1.5%, the parameter drift event is considered valid.

7. The method according to claim 1, characterized in that, When the parameter drift event is determined to be valid, the load models in the preset scenario-based candidate load model library are updated based on preset rules, and the updated load model is selected as the optimal load model for the corresponding scenario, including: A pre-built library of scenario-based candidate load models containing at least several typical scenarios; Starting from the current moment, each scenario-based candidate load model is substituted into the multi-business time series data within the future backtesting window for quantitative backtesting, and the root mean square error (RMS_ERR_i) of each scenario-based candidate load model within the backtesting window is calculated. The Reward_i score for the i-th scenario-based candidate load model is calculated based on the backtest root mean square error RMS_ERR_i: Reward_i = 1 / (1+RMS_ERR_i); The scenario-based candidate load model with the highest Reward_i score is selected as the optimal load model.

8. The method according to claim 1, characterized in that, The method further includes: Record the optimal load model for each time period to generate a sequence of optimal load models for the whole year (96×365). Generate a visualization view of the optimal load model sequence; Generate an evolution analysis report of the load model according to a preset cycle.

9. The method according to claim 1, characterized in that, The method further includes: Based on new multi-business time-series data, the scenario-based candidate load model library is evaluated periodically, and scenario-based candidate load models with continuously declining performance are marked or deleted. New scenario-based candidate load models are generated through cluster analysis.

10. A load model selection system based on quantitative evaluation of spatiotemporal characteristic evolution laws, characterized in that, The system includes: An initial unit is used to acquire multi-business time-series data for load modeling and to construct the time-series features of the multi-business time-series data; The identification unit is used to perform natural band-based quantitative analysis on the multi-business time series data through the constructed dynamic envelope channel to identify parameter drift events; The construction unit is used to construct a comprehensive feature vector based on the multi-business time-series data, the time-series features, and parameter drift events; The judgment unit is used to obtain the prediction error of the comprehensive feature vector through the trained LightGBM regression model, and to judge the parameter drift event through the prediction error. The result unit is used to select the optimal load model from the preset scenario-based candidate load model library based on preset rules when the parameter drift event is determined to be valid, and to use the updated load model as the optimal load model for the corresponding scenario.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-9.

12. An electronic device, characterized in that, include: The computer-readable storage medium as described in claim 11; as well as One or more processors for executing a program in the computer-readable storage medium.

Images