Power load forecasting method based on micro-incremental shrinkage and hidden state transition model

Abstract

The application discloses a power load prediction method based on a micro-incremental shrinkage and hidden state transition model, and comprises the following steps: S1, acquiring a total load sequence of a target prediction area and a plurality of micro power consumption object load sequences, forming a current macro load continuation state according to the total load sequence, and forming a micro power consumption object load incremental record according to the plurality of micro power consumption object load sequences; S2, determining a load incremental occurrence position of each micro power consumption object in a current rolling window according to the micro power consumption object load incremental record, and determining a window dominant incremental direction according to the current macro load continuation state. The application solves the problem that in the power load prediction scene, when the macro total load is still in a continuation state, the existing method cannot effectively identify the macro load path mutation period through the change of the micro power consumption object same-direction load increment from dispersion to concentration.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting technology, and in particular to a power load forecasting method based on a micro-incremental contraction and hidden state transition model. Background Technology

[0002] Electricity load forecasting is a fundamental technical aspect of power system operation control and dispatching planning, primarily assessing changes in electricity demand across a target region at different sampling times. With variations in residential, commercial, industrial, and distributed energy consumption, the total regional load no longer exhibits only smooth cyclical fluctuations but is also influenced by the start-up and shutdown, concentrated increases and decreases, and synchronous changes in electricity consumption behavior of multiple micro-level electricity consumers during localized periods. The grid operator needs not only to obtain future load values ​​but also to identify whether the macro-level load path has shifted from a continuous state to a sudden change. For the target forecasting region, the total load sequence reflects macro-level trends, while the load sequences of multiple micro-level electricity consumers reflect local sources of increase. Correlation between changes in the location of micro-level load increases and transitions in the macro-level load path becomes a crucial technical aspect of load forecasting.

[0003] In the prior art, Chinese patent CN102426674A discloses a power system load forecasting method based on Markov chains. The method first collects multiple sets of historical load data, maps the load values ​​of the previous and current times according to their value ranges to state values ​​within a state set, then calculates the state transition distance between adjacent times to form a state transition table; subsequently, it statistically analyzes the probability of each state transition distance, using the state transition distance with the highest probability as the basis for the transition at the time to be predicted, and determines the range of the load value at the next time moment based on the state of the load value at the previous time moment and the maximum probability transition distance. The paper "Short term load forecasting with discrete state Hidden Markov Models" proposes a discrete state Hidden Markov model method for hourly short-term load active power forecasting. It uses load data from a New York independent system operator from 2014 to 2017 and identifies features that can explain load power changes from weather, market, and calendar data. Considering the strong seasonality of the load sequence, the method sets up filtering processing and finally uses twenty-four discrete state Hidden Markov models with many states to perform forecasting, while comparing the forecast results with those of the operator and benchmark methods. The above schemes all revolve around load time-series state division, state transition probability, and hidden state inference to establish a prediction process.

[0004] Existing technical solutions primarily establish the transition relationship from the observed sequence to the prediction result based on the historical state of the total load, external characteristics, or discrete hidden states. The load increments of micro-level electricity users typically do not enter the state transition process as an independent constraint source capable of altering the degree of retention of the macro-level path state. When the total load in the target prediction area still exhibits a continuing upward trend, a continuing downward trend, or a stable trend, the locations of the same-direction load increments of multiple micro-level electricity users may have changed from dispersed to concentrated. The state transition probability at the macro-level sequence level easily incorporates this concentrated change into ordinary fluctuations. Existing solutions lack the construction of the contraction amount of the cross-window load increment location for same-direction increment objects, and also lack a state transition structure that embeds the contraction amount of the micro-increment location into the slow hidden state retention ratio generation process. This makes it difficult to directly determine the abrupt change period of the macro-level load continuation path under the appearance of total load continuation. Summary of the Invention

[0005] This invention provides a power load forecasting method based on micro-incremental contraction and hidden state transition model. It solves the technical problem of how to utilize the cross-window change in the location of load increments of multiple micro-electricity objects from dispersed to concentrated in the power load forecasting scenario of the target forecasting area, constrain the state preservation process of the current macro load continuation state in the hidden state transition model, so that the path change period generated when the macro load path maintains the continuation appearance can be determined and used as the output of the power load forecasting result.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a power load forecasting method based on a micro-incremental contraction and hidden state transition model, comprising the following steps: S1. Obtain the total load sequence and multiple micro-level electricity consumption object load sequences of the target prediction area, form the current macro-level load continuity status based on the total load sequence, and form the micro-level electricity consumption object load increment record based on the multiple micro-level electricity consumption object load sequences. S2. Based on the load increment records of micro-level electricity users, determine the location of the load increment of each micro-level electricity user within the current rolling window, and determine the dominant increment direction of the window based on the current macro-level load continuity status. S3. Based on the location of the load increment and the dominant increment direction of the window, determine the same-direction increment object, and generate the micro increment location contraction amount based on the load increment location of the same-direction increment object in the current scrolling window and the previous scrolling window. S4. Construct a hidden state transition model, taking the current macroscopic load continuation state as the slow hidden state, taking the microscopic electricity object load increment record as the fast hidden state, and embedding the shrinkage amount of the microscopic increment location into the state retention ratio generation process of the slow hidden state transition operator. S5. Based on the state retention ratio after the shrinkage of the embedded micro-increment, the slow hidden state is transferred to generate the macro load path state after the shrinkage constraint. S6. Determine the load path abrupt change based on the macro load path state after the contraction constraint. When the macro load path state after the contraction constraint meets the load path abrupt change condition, generate the macro load continuation path abrupt change period. S7. Outputs power load forecast results including the period of abrupt change in the macro load continuation path and the state of the macro load path after contraction constraints.

[0007] Preferably, step S1 specifically includes: The total load sequence is sampled according to the current rolling window and the previous rolling window to form a macroscopic window sequence; the load sequences of multiple microscopic electricity users are time-aligned according to the sampling time of the macroscopic window sequence to form a microscopic electricity user aligned load sequence. Calculate the direction of total load change between adjacent sampling times within the current rolling window and generate a cumulative result. Compare the cumulative result with a preset threshold to determine the current macroscopic load continuation status. Based on the aligned load sequence of micro-electricity users, the direction and magnitude of load increment for each micro-electricity user between adjacent sampling times are calculated to form a load increment record for each micro-electricity user.

[0008] Preferably, step S2 specifically includes: Extract the load increment direction and load increment magnitude of each micro-electricity object at each sampling moment within the current rolling window from the load increment record of each micro-electricity object, and form the current window increment time sequence; The load increment magnitude is compared with the preset candidate increment magnitude threshold to obtain a set of candidate increment locations for each micro-electricity consumer; the load increment direction within the candidate increment location set is screened for consistency to generate the load increment location of each micro-electricity consumer within the current scrolling window; The load path continuation direction is obtained by matching the current macro load continuation status, and this direction is determined as the window-dominant incremental direction.

[0009] Preferably, step S3 specifically includes: Extract the load increment direction corresponding to the location where the load increment occurs for each micro-level electricity consumer and form a location-direction correspondence; match the load increment direction of each micro-level electricity consumer with the dominant increment direction of the window, calculate the proportion of consistent directions, and filter out the same-direction increment objects and the same-direction location set; Get the position of the same direction load increment of the same object in the previous scroll window, and perform sequence matching between the position of the same direction load increment in the current scroll window and the previous scroll window to form cross-window same direction position pairs. Based on the cross-window same-direction position pairs, determine the position contraction reference span of the previous scrolling window and the current window position span of the current scrolling window respectively, and compare them to obtain the position span reduction amount; compare the position span reduction amount with the preset contraction judgment threshold, and when the condition is met, generate a micro increment based on the position span reduction amount to generate the position contraction amount.

[0010] Preferably, step S4 specifically includes: Construct a hidden state transition model that includes slow hidden state, fast hidden state, and slow hidden state transition operator, and limit the calculation process in the slow hidden state transition operator used to determine the degree of preservation of slow hidden state to the state preservation ratio generation process. The current macro load continuity state is encoded to form a slow-concealed state that carries the macro load path tendency; the load increment records of micro-electricity objects are incrementally encoded to form a fast-concealed state that carries the local load change information of the current rolling window. The basic state retention ratio is calculated based on the slow hidden state and the fast hidden state; the shrinkage amount at the location where the micro increment occurs is input into the state retention ratio generation process to generate a shrinkage constraint amount for the basic state retention ratio; the basic state retention ratio is updated based on the shrinkage constraint amount to obtain the state retention ratio after embedding the shrinkage amount at the location where the micro increment occurs and is written into the slow hidden state transition operator.

[0011] Preferably, step S5 specifically includes: The state transition control quantity is generated based on the state retention ratio after the shrinkage of the embedded micro-increment location; the path continuation component in the slow hidden state is retained based on the state transition control quantity to form the slow hidden state retention component. Combine the fast hidden state and the slow hidden state to perform candidate write calculations to form a candidate slow hidden state; determine the state write ratio according to the state transition control quantity, and perform write calculations on the path change components in the candidate slow hidden state to form the slow hidden state write component. By combining the slow hidden state retention component and the slow hidden state write component, a slow hidden state after contraction constraint is generated. The path continuation component and path change component are extracted to form a macro load path candidate state. After reading the path state, a macro load path state after contraction constraint is generated.

[0012] Preferably, step S6 specifically includes: Extract the path continuation component and the path change component from the macroscopic load path state after contraction constraints, and calculate the path state difference. The deviation of the path change component from the path continuation component is calculated based on the path state difference, forming the load path abrupt change judgment quantity; The load path mutation judgment quantity is compared with the preset threshold in the load path mutation condition. When the mutation condition is met, the macro load continuity path mutation period is generated according to the start and end sampling times of the current rolling window.

[0013] Preferably, step S7 specifically includes: Based on the start and end sampling times of the abrupt change period of the macro load continuation path, the corresponding path state segments are matched and extracted from the macro load path state after the contraction constraint to form the path state of the abrupt change period. The output range is constrained based on the path state during the abrupt change period, resulting in path state output content that includes both path continuation and path change components within the abrupt change period. By combining the path status output with the period of abrupt changes in the macro load continuation path, the power load forecast results are generated and output.

[0014] Preferably, before forming cross-window same-direction position pairs, it is first determined whether the same-direction incremental objects have load increment occurrence positions consistent with the dominant increment direction of the window in both the current scrolling window and the previous scrolling window. Only same-direction incremental objects that have same-direction load increment occurrence positions in both windows are used to form cross-window same-direction position pairs. The load increment occurrence positions participating in the calculation of position shrinkage reference span, position span reduction, and micro increment occurrence position shrinkage are all limited to load increment occurrence positions consistent with the dominant increment direction of the window.

[0015] Preferably, when embedding the shrinkage amount at the location of the micro-increment into the state retention ratio generation process, the slow hidden state and the fast hidden state are kept as the state inputs of the slow hidden state transition operator. The shrinkage amount at the location of the micro-increment is only used to generate the shrinkage constraint amount to update the basic state retention ratio. The retention degree of the slow hidden state is constrained by the state retention ratio generation process.

[0016] Preferably, when determining the slow hidden state retention component and the slow hidden state write component based on the state transition control quantity, when the state transition control quantity indicates a reduction in the retention range of the original macroscopic load path tendency of the slow hidden state during state transition, the slow hidden state retention component is reduced, and the slow hidden state write component is increased according to the state write ratio. The slow hidden state after contraction constraint is generated based on the reduced slow hidden state retention component and the increased slow hidden state write component, so that the macroscopic load path state after contraction constraint inherits the constraint result of the microscopic increment occurrence location contraction amount on the slow hidden state.

[0017] Preferably, the load path abrupt change condition includes the path continuation component corresponding to the current macro load continuation state, and the load path abrupt change determination quantity satisfying the preset threshold in the load path abrupt change condition. When the path continuation component corresponds to the current macro load continuation state and the load path abrupt change determination quantity satisfies the preset threshold in the load path abrupt change condition, it is determined that the macro load path state after the contraction constraint satisfies the load path abrupt change condition, and a macro load continuation path abrupt change period is generated according to the start and end sampling times of the current rolling window.

[0018] This invention provides a power load forecasting method based on a micro-incremental contraction and hidden state transition model, which has the following beneficial effects: 1. This invention proposes an improved load path determination method for hidden state transition models. It embeds the shrinkage amount of the micro-increment location into the state retention ratio generation process of the slow hidden state transition operator, transforming the dispersed load increments of micro-level electricity users into concentrated cross-window changes that directly affect the retention degree of the current macro-level load continuation state. Compared to existing methods that rely solely on historical total load states, state transition probabilities, or external characteristics for load forecasting, this invention does not treat micro-load changes as ordinary input features for parallel fusion. Instead, it limits the shrinkage amount of the micro-increment location as a constraint source for the state retention ratio. This reduces the inertial retention of the original macro-level load path tendency during state transitions in slow hidden states and increases the proportion of path change information written within the current rolling window. Consequently, even when the macro-level load path still exhibits a continuation state, it can form readable path change components.

[0019] 2. This invention proposes a load path abrupt change identification method for concentrated changes in the same direction of micro-level electricity users. It determines the location of load increments based on the load increment records of micro-level electricity users, and determines the dominant increment direction of the window by combining this with the current macro-level load continuity state. Then, it filters objects with the same direction of increments, compares the span of the load increment locations of objects with the same direction of increments in the current rolling window with that in the previous rolling window, and generates the shrinkage amount of the micro-level increment location. Compared to methods that only determine the upward, downward, or stable trend from the macro-level total load curve, this invention transforms the synchronous concentrated changes of multiple micro-level electricity users in time location into quantitative constraints that can participate in the implicit state transition. This preserves the source of local load changes under the appearance of macro-level load continuity, reduces the possibility of incorporating concentrated changes in the same direction of increments into ordinary load fluctuations, and helps to determine the period of abrupt changes in the macro-level load continuity path.

[0020] 3. This invention proposes a power load forecasting method comprised of a total load sequence, a micro-level load sequence of electricity consumers, a hidden state transition model, and path abrupt change determination. This method first generates the current macro-level load continuity state and micro-level load increment records, then generates the shrinkage amount at the location of the micro-level increment, and writes this shrinkage amount into the state retention ratio generation process. Subsequently, it generates the macro-level load path state after shrinkage constraints, and finally generates the macro-level load continuity path abrupt change period based on the deviation between the path continuity component and the path change component. Compared to forecasting schemes that directly output future load power values, this invention outputs power load forecasting results including macro-level load continuity path abrupt change periods and macro-level load path states after shrinkage constraints, giving the target forecast area a clear state source and time boundary for the load path level change process. Attached Figure Description

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the present invention; Figure 2 This is a flowchart of the load sequence acquisition and incremental record formation process of the present invention; Figure 3 This is a flowchart illustrating the determination of the load increment location and the dominant increment direction of the window in this invention. Figure 4 This is a flowchart illustrating the process of generating the shrinkage amount at the location where the micro-increment occurs according to the present invention. Figure 5 This is a flowchart illustrating the construction of the hidden state transition model and the embedding of the state retention ratio in this invention. Figure 6 This is a flowchart illustrating the generation of the macroscopic load path state after contraction constraint according to the present invention. Figure 7 The flowchart for generating abrupt changes in the macroscopic load continuation path is provided in this invention; Figure 8 This is a schematic diagram illustrating the positional contraction of the load increment in the forward and backward rolling windows according to the present invention. Figure 9 This is a schematic diagram of the incremental screening and state transition structure of the present invention; Figure 10 This is a schematic diagram comparing the path continuation component and the path change component of the present invention; Figure 11 This is a schematic diagram illustrating the formation of macroscopic states and microscopic incremental inputs in this invention; Figure 12 This invention provides a multi-plot diagram of the load of ten microscopic power-consuming objects. Figure 13 This is a biline chart of the rolling window metrics for the test set of this invention. Detailed Implementation

[0022] Example 1, such as Figure 1 As shown, a power load forecasting method based on a micro-incremental contraction and hidden state transition model includes: S1. Obtain the total load sequence and multiple micro-level electricity consumption object load sequences of the target prediction area, form the current macro-level load continuity status based on the total load sequence, and form the micro-level electricity consumption object load increment record based on the multiple micro-level electricity consumption object load sequences. S2. Based on the load increment records of micro-level electricity users, determine the location of the load increment of each micro-level electricity user within the current rolling window, and determine the dominant increment direction of the window based on the current macro-level load continuity status. S3. Based on the location of the load increment and the dominant increment direction of the window, determine the same-direction increment object, and generate the micro increment location contraction amount based on the load increment location of the same-direction increment object in the current scrolling window and the previous scrolling window. S4. Construct a hidden state transition model, taking the current macroscopic load continuation state as the slow hidden state, taking the microscopic electricity object load increment record as the fast hidden state, and embedding the shrinkage amount of the microscopic increment location into the state retention ratio generation process of the slow hidden state transition operator. S5. Based on the state retention ratio after the shrinkage of the embedded micro-increment, the slow hidden state is transferred to generate the macro load path state after the shrinkage constraint. S6. Determine the load path abrupt change based on the macro load path state after the contraction constraint. When the macro load path state after the contraction constraint meets the load path abrupt change condition, generate the macro load continuation path abrupt change period. S7. Outputs power load forecast results including the period of abrupt change in the macro load continuation path and the state of the macro load path after contraction constraints.

[0023] like Figure 2 As shown, step S1 specifically involves: Obtain the total load sequence for the target prediction region and represent it as a timestamped sequence. , This indicates the total load sampling time for the target prediction area. This represents the total load value of the target prediction area at the corresponding total load sampling time. Indicates the total number of load sampling points; Set a uniform sampling period , based on the starting sampling time Construct a unified index axis ,index , Indicates the number of sampling points on the uniform index axis, and sets the window length. and the starting index of the current scroll window ,satisfy and ; The current scrolling window is represented as The previous scrolling window is represented as For each Select from the total load sequence of the target prediction area ,when At that time, Mapped to , Indicates the sampling time matching tolerance, when At that time, select the location in chronological order. Linear interpolation is performed on the effective total load samples from both sides, and the interpolated total load value is mapped to... ,Depend on Load value arrangement Form a macroscopic window sequence; Obtain the load sequence of multiple micro-level electricity users, and then... The load sequence of a micro-level electricity consumer is represented as a timestamped sequence. , Indicates the index of micro-level electricity-consuming objects. Indicates the number of micro-level electricity users. Indicates the first Load sampling time of each micro-level electricity consumer Indicates the first The load value of each micro-level electricity consumer at the corresponding load sampling time. Indicates the first The number of sampling points for each micro-level electricity user; For each And each In the Select from the load sequence of individual micro-electricity users ,when At that time, Mapped to ,when At that time, select the same micro-electricity consumer in chronological order. The effective load sample values ​​on both sides are linearly interpolated, and the interpolated load values ​​are mapped to... ,when When located outside the load sequence boundary of the same micro-level electricity consumer and unable to obtain effective load sampling values ​​on both sides, the distance will be... The most recent effective load sample value is mapped to ,Depend on and Load value arrangement Forming a load sequence aligned with micro-level electricity users; The direction of total load change between adjacent sampling times within the current rolling window is calculated based on the macroscopic window sequence, with preset thresholds including a total load amplitude threshold. Cumulative percentage threshold and consecutive number threshold For each ,Compare and Differences in load values: When the load difference is greater than the total load amplitude threshold At that time, the direction of total load change It has been determined to be in an upward direction; When the load difference is less than the negative total load amplitude threshold At that time, the direction of total load change If the load difference is determined to be in the downward direction, it means the load difference is within the total load amplitude threshold. Within the defined range, the direction of total load change If the direction is determined to be stable, the maximum consecutive occurrences and percentages of the upward, downward, and stable directions within the current scrolling window are compiled as follows: , , , These represent the maximum consecutive occurrences of the upward, downward, and level-off directions within the current scrolling window, respectively. , , These vectors represent the proportions of the upward, downward, and level directions within the current scrolling window. As a cumulative result of the direction of change in total load; The cumulative result of the total load change direction is compared with a preset threshold, and the current macro load continuity status is determined based on the comparison result. Current macroeconomic load status The possible values ​​include continuation of the upward trend, continuation of the downward trend, and continuation of the stable trend: when Reaching the threshold for consecutive counts ,or Reaching the cumulative percentage threshold At that time, the current macroeconomic load continuation status will be determined as an upward continuation; when Reaching the threshold for consecutive counts ,or Reaching the cumulative percentage threshold At that time, the current macroeconomic load status will be determined as a continuation of the decline; when Reaching the threshold for consecutive counts ,or Reaching the cumulative percentage threshold At that time, the current macroeconomic load status will be determined as stable continuation; If multiple directions simultaneously meet the preset threshold, the current macro load continuation state corresponding to the direction with the highest proportion is selected. If multiple directions have the same proportion, the sampling time at the end of the current rolling window is selected. The current macro-load continuity state corresponding to the direction of total load change; Based on the aligned load sequence of micro-level electricity users, the direction and magnitude of load increment for each micro-level electricity user between adjacent sampling times are calculated. The preset threshold also includes a micro-load magnitude threshold. For each And each To ensure that the number of incremental records in the current scrolling window is equal to the number of records in the current scrolling window... To keep all sampling times consistent, the load increment direction corresponding to the starting sampling time of the current rolling window is set to a stable direction, and the corresponding load increment amplitude is set to 0. For each sampling time within the current rolling window, excluding the initial sampling time, the load increment direction and load increment magnitude are generated based on the load value difference between the current sampling time and the previous sampling time. Therefore, each sampling moment within the current rolling window corresponds to a load increment direction and a load increment magnitude; Compare and The load value difference is determined, and the absolute magnitude of the load value difference is defined as the load increment amplitude. : When the load difference is greater than the micro load amplitude threshold At that time, the direction of the load increment It has been determined to be in an upward direction; When the load difference is less than the negative micro load amplitude threshold At that time, the direction of the load increment The direction of the decline has been determined. When the load difference is within the micro load amplitude threshold Within the defined range, the direction of load increment. Determined to be a stable direction; The first At the sampling time, a micro-level electricity consumer The corresponding record is represented as Each To create a two-dimensional record that includes both the direction and magnitude of the load increment, the load increment record for a micro-level electricity consumer is written as follows: The load increment records of micro-level electricity users are arranged according to the micro-level electricity user index and the sampling time index. The input organization for the hidden state transition model is a one-dimensional discrete state label. and scale Two-dimensional record sequence .

[0024] like Figure 3 As shown, step S2 specifically involves: According to the load increment records of micro-level electricity users Extract the load increment direction and load increment magnitude at each sampling moment within the current scrolling window. For each micro-level electricity consumer, within the current scrolling window... Read in ascending order according to the sampling time index ,Will The load increment direction field in the data is used as ,Will The load increment magnitude field in the data is used as And record and represent a single sampling time as , Indicates the first The sampling time of each micro-electricity consumer within the current scrolling window The corresponding load increment record is... All Arranged according to the sampling time index, forming the [number]th [item]. Current window increment timing of a micro-level electricity consumer ; The current window incremental time series is organized according to the micro-electricity consumption object index and the sampling time index. Each record contains two fields: load increment direction and load increment magnitude. The load increment direction can take the values ​​of rising direction, falling direction, and leveling off direction. The candidate increment position set is determined based on the current window increment time sequence, and a candidate increment amplitude threshold is set. , This represents the preset threshold used to determine whether the magnitude of a load increment enters the candidate increment location set; the candidate increment magnitude threshold. Not less than the micro load amplitude threshold For each micro-level electricity user Read the incremental timing sequence of the current window one by one. The magnitude of the load increment at each sampling time The magnitude of the load increase With candidate increment magnitude threshold Comparison: When the load increment magnitude Reaching the candidate increment magnitude threshold At that time, the sampling time index will be used. This was identified as a candidate incremental position. When the load increment magnitude The candidate increment threshold was not reached. At that time, the sampling time is not indexed. This was identified as a candidate incremental position. No. The set of candidate incremental locations for each micro-electricity consumer is represented as follows: , Indicates the first The set of sampling times obtained by comparing the load increment amplitude of a micro-level electricity consumer within the current rolling window. If there is no candidate increment amplitude threshold within the current rolling window... At the sampling time, the first The candidate incremental location set for each micro-electricity user is an empty set; Based on the current window increment time series, load increment direction within the candidate increment location set is screened for consistency, and a threshold for the number of directions with consistency is set. Percentage threshold for direction consistency , This represents the minimum number of sampling times required for direction-consistent filtering. This represents the minimum proportion required for directional consistency filtering, for each non-empty set of candidate incremental positions. Incremental timing from the current window Read the load increment direction corresponding to each candidate increment position. The upward, downward, and stable directions are respectively counted in the set of candidate increment positions. The load increment direction with the highest occurrence count is determined as the load increment direction for consistency screening. , Indicates the first The direction reference used for determining the consistency of direction for each micro-level electricity consumer within the current scrolling window; if the number of occurrences of multiple load increment directions is the same, then the candidate increment location set is selected. The load increment direction is read in descending order according to the sampling time index, and the first load increment direction read that has the same number of occurrences is determined as... ; In determining Then, the set of candidate incremental positions is statistically analyzed. The direction of medium load increment is equal to The number of sampling times is calculated, and the number of sampling times is divided by the set of candidate incremental positions. The total number of sampling times is used to obtain the proportion of consistent directions. If the number of sampling times reaches the threshold for consistent directions... Or the proportion of directions in the same direction reaches the threshold for the proportion of directions in the same direction. Then determine the first Each micro-level electricity consumer satisfies the condition of consistent direction, and a set of candidate incremental locations is retained. The direction of medium load increment is the same as The same sampling time index forms the first Location of load increment for each micro-level electricity consumer within the current scrolling window , Indicates the first The set of sampling times of a micro-level electricity consumer within the current rolling window that meet both the load increment magnitude comparison condition and the load increment direction consistency condition. If the number of sampling times does not reach the threshold for the number of consistent directions. Furthermore, the proportion of directions aligned did not reach the threshold for the proportion of directions aligned. Then the first The location of the load increment of each micro-electricity consumer within the current scrolling window is determined to be an empty set; The dominant incremental direction of the window is determined based on the current macroeconomic load continuation status, and the current macroeconomic load continuation status is denoted as... The direction of load path continuation is denoted as , This indicates the direction of load change corresponding to the current rolling window, which is matched with the current macro load continuity status: When the current macroeconomic load continues When the load path continues to rise, the direction of load path continuation will be changed. It has been determined to be in an upward direction; When the current macroeconomic load continues When the load path continues to fall, the direction of load path continuation will be changed. The direction of the decline has been determined. When the current macroeconomic load continues To ensure smooth continuation, the load path continuation direction is... Determined to be a stable direction; Continuation direction of load path Determined as the window-dominant incremental direction , This indicates the main load change direction within the target prediction area in the current rolling window, and the dominant increment direction of the window. Location of load increment The corresponding load increment directions adopt the same direction value system, so that the load increment location of each micro-electricity object in the current rolling window can be matched according to the dominant increment direction of the window.

[0025] like Figure 4 As shown, step S3 specifically involves: Based on the location of load increment occurrence for each micro-level electricity consumer within the current rolling window and window-dominated incremental direction For each micro-electricity consumer object, the load increment direction corresponding to the location of each load increment is extracted according to the micro-electricity consumer object index. Read from the current window incremental time series Index of each sampling time within Corresponding load increment direction And bind the sampling time index and the load increment direction to , No. The positional and directional correspondence of individual micro-electricity-consuming objects is represented as follows: , This indicates the binding result between the location of the load increment and the corresponding direction of the load increment. If it is an empty set, then It is an empty set, and the first... Individual electricity users are not included in the calculation of the direction consistency ratio; Based on the correspondence of position and direction The load increment direction of each micro-level electricity user is compared with the window's dominant increment direction. Perform matching, and establish a correspondence between the non-empty positions and directions. The direction of the statistical load increment is equal to the direction of the window-dominant increment. The number of locations where load increments occur is calculated, and the statistically obtained number is divided by... The total number of locations where internal load increments occur is used to determine the proportion of locations with consistent direction. , Indicates the first The percentage of load increments occurring in the same direction as the dominant increment direction within the current scrolling window for each micro-electricity consumer, with preset thresholds including the same-direction ratio threshold. , This represents the minimum proportion of directional consistency used to filter incremental objects in the same direction. Same direction proportional threshold When comparing, Reaching the same direction ratio threshold At that time, the first Each micro-level electricity consumer is identified as a unidirectional incremental object, when The same direction ratio threshold was not reached. At that time, the first Each micro-level electricity consumer is identified as a unidirectional incremental object, and the indices of all unidirectional incremental objects are used to form a unidirectional incremental object index set. , This represents the set of indices of micro-electricity-consuming objects that meet the condition of proportional comparison based on direction consistency. For each From the location where the load increment occurs The direction of the retained load increment is equal to the direction of the window-dominant increment. The sampling time index forms a set of positions in the same direction. , Indicates the first The set of load increment locations for each same-direction increment object within the current scrolling window, aligned with the dominant increment direction of the window. To obtain the load increment location of the same same-direction increment object within the previous scrolling window, the data is read from the load increment records of the micro-level electricity users. corresponding The magnitude of the load increase With candidate increment magnitude threshold Compare and directionalize the load increment. With window-dominated incremental direction The matching will be performed to ensure that the load increment magnitude reaches the candidate increment magnitude threshold. And the direction of load increment is equal to the direction of window-dominant increment. The sampling time index forms the set of positions in the same direction within the previous scrolling window. , Indicates the first The set of locations where load increments occur that are consistent with the dominant increment direction of the window within the previous scrolling window; Before forming cross-window same-direction position pairs, determine Check if it is an empty set, and determine the result. Whether a set is empty depends only on whether it is empty. and When neither set is empty, the first set will be... A single incremental object in the same direction is used to form cross-window position pairs in the same direction, which will satisfy... and The indices of mutually exclusive incremental objects, which are all non-empty sets, form a set of matching mutually exclusive incremental object indices. , This represents a set of indices of load increment objects in the same direction as the dominant increment direction in both the current and previous scrolling windows. ,Will and The samples are sorted in ascending order by their sampling time indices, and matching is performed using the same sorting order. When the number of sampling times in the two sets is different, the matching count is determined by the smaller number of sampling times, and only the sampling time indices that have already been matched are retained. The cross-window co-directional position pairs of co-directional incremental objects are represented as follows: , This indicates the matching result between the location of the load increment in the previous scrolling window and the location of the load increment in the current scrolling window for the same incremental object in the same direction. This represents the matching index during the matching process with the same sort order. Indicates the number of the previous scrolling window The location where the matched load increment occurred. Indicates the number of elements in the current scrolling window. For each load increment location that has been matched, the load increment locations generated by cross-window same-direction location pairs, location contraction reference span, location span reduction, and micro-increment location contraction are all limited to load increment locations that are consistent with the dominant increment direction of the window. Based on cross-window same-direction position Determine the first The location where the initial load increment of each unidirectional incremental object has been matched in the previous scrolling window. and the location where the load increment occurs , This indicates the matching position with the smallest index at the sampling time within the previous scrolling window. This indicates the position of the maximum matching index within the previous scrolling window. and The number of sampling intervals between them is determined as the location shrinkage reference span. , This represents the distribution span of the locations where the same-direction load increments occur within the previous rolling window, with the dimension being the number of sampling intervals. It is based on the same-direction locations within the same span of the window. Determine the location of the initial load increment that has been matched within the current scrolling window. and the location where the load increment occurs , This indicates the matching position with the smallest index at the current sampling time within the scrolling window. This indicates the position of the largest matching index within the current scrolling window at the sampling time. and The number of sampling intervals between them is determined as the current window position span. , It represents the distribution span of the locations where the same-direction load increments occur within the current rolling window, with the dimension being the number of sampling intervals; span of the current position With positional contraction reference span The comparison generates the reduction in position span, and the preset threshold also includes a shrinkage determination threshold. , This represents the minimum reduction in the span of the location where the load increment occurs to determine whether it has contracted; its dimension is the number of sampling intervals. Greater than zero: When the position shrinks, the reference span Greater than the current position span ,and Compared to The reduction result reaches the shrinkage judgment threshold. At that time, the judgment of the first A number of incremental objects in the same direction satisfy the shrinkage determination condition; When the position shrinks, the reference span Not greater than the current position span ,or Compared to The reduction result did not reach the shrinkage determination threshold. At that time, the judgment of the first The same incremental objects in the same direction do not meet the shrinkage determination criteria; The location of the micro-increment is the amount of contraction. Generate based on the set of matching indices of incremental objects moving in the same direction, the current position span, the position contraction reference span, and the contraction determination criteria: ; in, This represents the amount of contraction at the location where the micro-increment occurs, with the dimension being the number of sampling intervals. This represents the set of indexes of objects that can be matched in the same direction of incremental growth. Indicates the index of micro-level electricity-consuming objects. This indicates an indicator function that takes the value when the condition within the parentheses is true. The value is taken as follows when the condition inside the parentheses is not true. , Indicates the first The positional shrinkage reference span of each incremental object in the same direction is measured in units of sampling intervals. Indicates the first The span of the current window position of each incremental object in the same direction, measured in the number of sampling intervals. This represents the shrinkage determination threshold, with the dimension being the number of sampling intervals. Indicates the first The reduction in the position span of each incremental object in the same direction, measured in units of sampling intervals. Indicates the first The number of load increment locations within the current scrolling window that are aligned with the dominant increment direction of the window, representing a dimensionless count. Indicates the first The number of load increment locations that align with the dominant increment direction of the window within the previous scrolling window for each co-directional increment object is a dimensionless count. The numerator of the formula is the cumulative value of the sampling interval after weighting the reduction in the location span of each co-directional increment object according to the number of load increment locations. The denominator is the cumulative value of the number of load increment locations participating in the weighting. Therefore, the shrinkage of micro-increment locations... Maintaining the same dimensions as the reduction in position span, the formula only includes load increments occurring at locations consistent with the dominant increment direction in both the current and previous scrolling windows, and whose reduction in position span meets the contraction criteria. When no load increments meeting the contraction criteria exist, the contraction amount at the location of the micro-increment is included in the calculation. The value is determined to be zero.

[0026] like Figure 5 As shown, step S4 specifically involves: Constructing a hidden state transition model , This represents the implicit state transition model used to form the macroscopic load path state. This represents the model parameters of the hidden state transition model. This includes parameters for the state encoding layer, incremental encoding layer, linear mapping parameters during the state retention ratio generation process, and candidate write linear mapping layer parameters, as well as model parameters. The training model is determined offline by training the historical total load sequence and the historical micro-level load sequence of electricity users according to the same window rules. The input fields in the training samples include the historical current macro-level load continuity state, the historical micro-level load increment record, and the contraction amount of the historical micro-level increment location. The training objective adopts the labeled macro-level load path state corresponding to the historical rolling window, and the hidden state transition model. Fixed model parameters during inference ; During offline training, the historical total load sequence and the historical micro-level electricity object load sequence are extracted according to the same window rules to generate the historical current macro-level load continuation status, historical micro-level electricity object load increment records, and historical micro-level increment occurrence location shrinkage amount. The labeled macro-level load path status corresponding to the historical rolling window is used as the training label. The hidden state transition model takes the historical current macro load continuation state, the historical micro load increment records of electricity users, and the shrinkage amount of the historical micro increment location as training inputs, and the labeled macro load path state as the training target. The model parameters are updated according to the difference between the macro load path state output by the model and the training label. When the preset number of training rounds is reached, or when the difference in the validation set does not decrease within the preset number of consecutive rounds, training is stopped and the model parameters are fixed. Hidden state transition model The system includes a state coding layer, an incremental coding layer, a slow concealment state, a fast concealment state, and a slow concealment state transition operator. The state coding layer is used to convert the current macroscopic load continuation state into a state expression acceptable to the slow concealment state. The incremental coding layer is used to convert the microscopic electricity object load increment record into an incremental expression acceptable to the fast concealment state. The slow concealment state transition operator is used to perform state transition on the slow concealment state according to the slow concealment state, the fast concealment state, and the state retention ratio. The sub-process within the slow concealment state transition operator used to determine the degree of slow concealment state retention is limited to the state retention ratio generation process, and the state retention ratio generation process is set inside the slow concealment state transition operator, so that the state retention ratio serves as the internal ratio basis of the slow concealment state transition operator. Continue the current macroeconomic load status As input to the state coding layer, the current macroscopic load continuation state The values ​​include rising continuation, falling continuation, and stable continuation. The state coding layer first assigns the current macroeconomic load continuation state to... Convert to a three-dimensional state label vector , The status code input represents the current macroeconomic load continuity status. , , These represent the flags corresponding to the continuation of the upward trend, the continuation of the downward trend, and the continuation of the stable trend, respectively. When the current macroeconomic load continuation status is the continuation of the upward trend... Set as The other two marker positions are When the current macroeconomic load continues to decline, Set as The other two marker positions are When the current macroeconomic load continues to be stable, Set as The other two marker positions are The state coding layer includes an input dimension of Output dimension is The linear mapping layer and the limiting layer, the linear mapping layer will convert the three-dimensional state identification vector Convert to The state vector is dimensional, and the amplitude limiting layer truncates components in the state vector that exceed a preset state amplitude range to within that range, forming a slow hidden state. , The dimension representing the slow hidden state. This represents the slow hidden state, which carries the macro load path tendency corresponding to the current macro load continuation state. The incremental load records of micro-level electricity consumers are used as input to the incremental encoding layer. The incremental encoding layer reads the data within the current scrolling window from the incremental load records of the micro-level electricity consumers. For each And each The incremental coding layer will direct the load increment direction Convert to 3D direction identifier vector and the magnitude of the load increment The amplitude field is concatenated with the three-dimensional direction identifier vector to form the incremental encoded input at a single sampling time. , Indicates the first At the sampling time, a micro-level electricity consumer The four-dimensional incremental encoding input, the incremental encoding layer divides all according to the micro-electricity consumption object index and the sampling time index. The organization is of a certain size Incremental input tensor; The object-dimensional convergence layer in the incremental coding layer at each sampling time right The four-dimensional incremental encoded inputs are aggregated into an eight-dimensional object aggregation vector. , Indicates the sampling time The object-dimensional aggregation result has eight fields, in order: average value of the upward direction indicator, average value of the downward direction indicator, average value of the stable direction indicator, average load increment magnitude, maximum value of the upward direction indicator, maximum value of the downward direction indicator, maximum value of the stable direction indicator, and maximum value of the load increment magnitude. The time-dimensional aggregation layer in the incremental encoding layer performs aggregation on the current scrolling window. Time field aggregation is performed, generating window mean, window maximum, and window end values ​​for each of the eight fields, and then concatenating them in field order to form a 24-dimensional time aggregation vector. , This represents the time convergence vector carrying the micro-load increment distribution within the current scrolling window. The linear mapping layer in the incremental coding layer converts the 24-dimensional time convergence vector... Mapped to Dimensional hidden state , This indicates the fast-hidden state of the local load change information within the current scrolling window corresponding to the load increment record of the micro-electricity user object; Within the slow hidden state transition operator, the slow hidden state is... and fast hidden state The state input is maintained for the state retention scaling generation process, which will then slowly conceal the state. and fast hidden state spliced ​​as The state-scale input vector is used to generate the basic state retention ratio through a two-layer linear mapping structure. The input dimension of the first-level linear mapping structure is The output dimension is This is used to combine slow hidden states and fast hidden states. The input dimension of the second-layer linear mapping structure is... The output dimension is This is used to form the dimension-by-dimensional retention ratio, and to limit each dimension to a specific value by truncation. to Between, the basic state retention ratio This indicates the degree to which the slow hidden state retains the original macroscopic load path tendency in the corresponding dimension before the micro-increment occurs and the shrinkage amount is embedded. The location of the micro-increment is reduced by the amount of shrinkage. The process of generating the input state retention ratio and setting the shrinkage threshold. , This represents a preset threshold used to determine whether the contraction amount at the location of the micro-increment participates in the retention ratio of the constraint base state. The unit is the number of sampling intervals. The contraction threshold is less than the maximum number of sampling intervals allowed by the current rolling window, ensuring that the number of sampling intervals in the calculation of the contraction constraint amount is positive and satisfies... The location of the micro-increment will shrink. With shrinkage threshold By comparing, a state is formed in which the comparison conditions are met. ,when Reaching the shrinkage threshold When comparing the satisfaction status of the conditions. Set to the satisfied state, when The contraction threshold was not reached. When comparing the satisfied states of the comparison conditions. Set to an unsatisfied state, the location of the micro-increment shrinkage occurs. Used only for forming the retention ratio for the basic state The contraction constraint is not considered as a slow hidden state. The input content is not considered as a hidden state. The input content; Based on the retention ratio of the basic state according to the shrinkage constraint The update is performed to determine the state retention ratio after the shrinkage at the location where the embedded micro-increment occurs. During the update, the basic state retention ratio is processed dimension by dimension according to the slow-hiding state, and the value of each updated dimension is limited to the lower limit of the preset ratio. and preset ratio upper limit Between, and satisfy , This represents the minimum permissible percentage of the slow concealment state that should be preserved. This represents the maximum permissible proportion of the slow-hidden state retention, the proportion of the state retained after the shrinkage at the location of the embedded micro-increment. Determine as follows: ; ; in, This indicates the proportion of the state retained after the location of the embedded micro-increment shrinkage. The Dimensional value, This indicates the proportion of the state retained after the location of the embedded micro-increment shrinkage. The dimension index representing the slow hidden state. The dimension representing the slow hidden state. This indicates the maximum permissible percentage of the slow concealment state that can be preserved. This represents the minimum permissible percentage of the slow concealment state that should be preserved. Indicates the percentage of basic state retention The Dimensional value, Indicates the percentage of the base state that is retained. This indicates an indicator function that takes the value when the condition within the parentheses is true. The value is taken as follows when the condition inside the parentheses is not true. , This represents the amount of contraction at the location where the micro-increment occurs, with the dimension being the number of sampling intervals. This represents the shrinkage threshold, with the dimension being the number of sampling intervals. This indicates the current length of the scrolling window. This indicates the maximum number of sampling intervals allowed in the current scrolling window. This indicates taking the smaller value. This indicates taking the larger value. This indicates that the smaller value is taken between the contraction amount at the location where the micro-increment occurs and the maximum number of sampling intervals allowed by the current rolling window. This represents the range of sampling intervals from the shrinkage threshold to the maximum number of sampling intervals allowed by the current scroll window. Represents the contraction constraint amount The Dimensional value, Indicates the percentage of the base state that is retained. The contraction constraint is given by the fractional term in the formula, which is a dimensionless proportion. , , , and All are dimensionless proportional values; When the comparison condition is satisfied To satisfy the state, the contraction constraint amount Location contraction occurs with micro-increment Relative to the shrinkage threshold It increases with the increase of the base state, and retains the proportion of the base state. Perform a dimensional downsizing, when the comparison conditions are met. When the state is not satisfied, the indicator function takes the value of contraction constraint amount Each dimension does not retain the proportion of the basic state The proportion of the state retained after the reduction in the location of the embedded micro-increment occurs. for The dimension-wise proportional vector is used to represent the degree to which the slow hidden state retains the original macroscopic load path tendency during state transition; The proportion of the state retained after the shrinkage at the location of the embedded micro-increment. Write the slow-hidden state transition operator, and the slow-hidden state transition operator reads the slow-hidden state during state transition. Quick Hidden State The proportion of the state retained after the location of the embedded micro-increment shrinkage. and will As the basis for the state transition of the slow hidden state, this structure allows the state retention ratio generation process to retain the slow hidden state and the fast hidden state as state inputs, and limits the contraction amount of the micro increment occurrence location as the constraint source of the state retention ratio, so that the contraction amount of the micro increment occurrence location directly constrains the degree of retention of the slow hidden state through the contraction constraint amount.

[0027] like Figure 6 As shown, step S5 specifically involves: The proportion of the state retained after the shrinkage at the location of the embedded micro-increment. Input a slow hidden state transition operator, embedding the state retention ratio after the shrinkage amount at the location of the micro-increment. for dimensional proportional vector, slow hidden state transition operator pair Perform dimensional validation to make Dimensions and slow hidden states If the dimensions are consistent, and the dimension verification passes, the slow hidden state transition operator will... Each dimension value in the code is copied into the state transition control variable in the same dimensional order. , This represents the state transition control variable that acts on the preservation and writing of the slow hidden state. Indicates the first The reserved control values ​​for each slow hidden state dimension, with a value range of [value range missing]. to , Dimension index representing the slow hidden state, state transition control variable State retention ratio after the location shrinkage of the embedded micro-increment Dimensional correspondence enables the slow hidden state transition operator to perform state transitions according to the proportion of positional contraction constraint after the micro-increment occurs; Based on state transition control quantity Determine the slow concealment state During state transition, a retention range threshold is set for the retention range of the original macroscopic load path tendency. , This represents the proportional threshold used to define the range of slow hidden states to be preserved, and its value range is [value range missing]. to , state transition control quantity Each dimension With retention range threshold Comparison will satisfy The set of dimensions is denoted as , The dimension range that represents the preservation of the original macroscopic load path tendency during state transition in a slowly hidden state will satisfy [the following condition]. The set of dimensions is denoted as , This indicates that during the slow-concealing state, the scope of the path change information written within the current scrolling window is increased during state transition, and a preset lower limit is set. , This represents the minimum permissible percentage for the slow-concealed state to be preserved. China belongs to The components, according to the corresponding Preserve the slow-hidden state components, for the slow-hidden state China belongs to The amount, according to the preset lower limit Preserve the slow-hidden state components to form slow-hidden state preserved components. , This represents the component formed after the path continuation component carrying the current macroscopic load continuation state in the slow hidden state is retained and calculated. According to the fast concealment state and slow concealment state Within the slow hidden state transition operator, candidate write computation is performed. The candidate write computation first converts the fast hidden state... and slow concealment state Concatenate the candidate input vectors according to their dimensions. , express Candidates are written into the input vector. The candidate write-to-linear mapping layer is entered, and the candidate write-to-linear mapping layer uses a weight matrix. and bias vector , Indicates having lines and Column weight matrix, express 3D bias vector, weight matrix and bias vector Model parameters belonging to the hidden state transition model The candidate is written into the linear mapping layer, which reads the weight matrix for each output dimension. The corresponding row in the middle Each weight, will Each weight and candidate input vector In Each field is multiplied and summed, then the bias vector is added. The bias value of the corresponding dimension is used to set the upper limit of the state amplitude. , This represents the maximum allowed absolute magnitude of the slow hidden state component. The output of the candidate written linear mapping layer undergoes state magnitude limiting processing to form the candidate slow hidden state. , This indicates a candidate state that can be used to update the slow-concealing state. The candidate slow-concealing state carries information about the local load changes within the current scrolling window introduced by the fast-concealing state. Based on state transition control quantity Determined and slow-concealed state preserved components Corresponding state write ratio , This indicates the level of writing to candidate slow hidden states. dimensional scaling vector Indicates the first Write control values ​​for each slow hidden state dimension, for each dimension ,when At that time, write the state to the ratio. Determined as minus The value after that, when At that time, write the state to the ratio. Determined as minus The subsequent values ​​are written according to the state ratio. For candidate slow hidden states Perform write calculations to form slow hidden state write components. , This represents the path change component from writing slow hidden states to candidate slow hidden states according to the state write ratio; Components are preserved based on the slow concealment state. and slow hidden state write component For slow hidden state The state transition is performed, and the core computational object of the state transition is the slowly hidden state after the contraction constraint. Because of the slow hidden state after the contraction constraint Simultaneously, it undertakes the constraints of the state transition control quantity on the degree of preservation of the slow hidden state and the writing of local load change information by the candidate slow hidden state, shrinking the slow hidden state after constraint. Each dimension is generated according to the following formula: ; ; in, Represents the slow hidden state after contraction constraint. The Dimensional value, This represents the slow hidden state after the contraction constraint. This indicates state amplitude limiting processing; when the value in parentheses is higher than... Time to take When the value in parentheses is lower than Time to take When the value inside the parentheses is located Keep the value inside the parentheses. This represents the maximum absolute amplitude allowed for the slow concealment state component. This indicates an indicator function that takes the value when the condition within the parentheses is true. The value is taken as follows when the condition inside the parentheses is not true. , Represents state transition control quantity The Maintain control values. This represents the state transition control variable. Indicates the threshold for the retention range. This represents the minimum permissible percentage of the slow concealment state that should be preserved. Indicates slow concealment state The Dimensional value, This indicates a slow-concealing state. Indicates candidate slow hidden state The Dimensional value, This indicates a candidate slow hidden state. The dimension index representing the slow hidden state. The dimension representing the slow hidden state is shown in the formula. The bracketed terms of the multiplication are the components preserved in the slow hidden state at the 1st digit. The retention ratio adopted by the dimension, and The parenthesized terms in the multiplication are the components written to the slow hidden state in the 1st... The state write ratio used by the dimension , , Both the indicator function and the value of the indicator function are dimensionless proportional values. , , and All are hidden state values; When the state transition control quantity When indicating a reduction in the extent to which the original macroscopic load path tendency is preserved during state transitions in a slow concealed state, the following conditions must be met: The dimension is based on the preset ratio lower limit. right To retain, and in accordance with right Perform the write operation to meet the requirements. Dimensions according to right To retain, and in accordance with right Write to the slow hidden state after shrinking constraints. The constraint results of the contraction amount at the location of the micro-increment on the degree of preservation of the slow hidden state; Based on the slow concealment state after contraction constraint Extract path continuation components from the component source relations preserved by the slow hidden state transition operator. and path change components , This indicates the path continuation component that carries over the current state of macroeconomic load. The first component representing the path continuation Dimensional value, This indicates the path change component that receives the path change information within the currently scrolling window. The first component representing the path change Dimensional values, path continuation components The path change component is composed of the numerical values ​​corresponding to the slow hidden state preserved components in the formula. Composed of the values ​​corresponding to the slow-concealing state write components in the formula, and Arranged according to a fixed field order, forming a macroscopic load path candidate state. , This represents the candidate path states used to generate the macroscopic load path states after contraction constraints, with dimensions of... , among which the former Each field represents a path continuation component, followed by... Each field represents a path variation component; Based on state transition control quantity Macroscopic load path candidate status Perform path status reading, which will include path continuation components. China satisfies The dimension is preserved as the path continuation component after reading. and the path continuation component China satisfies The dimension is set to zero, and the path status reading will include path change components. China satisfies The dimension is preserved as the path change component after reading. and path change components China satisfies The dimension is set to zero. This indicates the continuation of the path after reading, with dimensions of [missing information]. , This represents the path changes after reading, with the dimension being... Continue the read path components and the path change components after reading Generate the macroscopic load path state after contraction constraints by arranging the fields in a fixed order. , This represents the macroscopic load path state formed after the state retention ratio, constrained by the location of the micro-increment, is applied to the slow hidden state; the dimension is... , among which the former Each field represents the continuation of the read path. Each field represents the path change components after reading.

[0028] Step S6 is as follows: Macroscopic load path state after contraction constraint Current macroeconomic load status and the current scrolling window The sampling time is the input object, and the macroscopic load path state after contraction constraints. The dimension is ,forward Each field represents the continuation component of the read path. ,back Each field represents the path change components after reading. , This indicates that the path continuation component after reading is in the first position. The values ​​can be taken in the dimension of each slow hidden state. This indicates that the path change component after reading is in the first position. The values ​​can be taken in the dimension of each slow hidden state. The dimension index representing the slow hidden state; From the macroscopic load path state after contraction constraints Extracting path continuation components after reading according to a fixed field order and the path change components after reading The read path continues the component Based on the path continuity information corresponding to the current macroeconomic load continuity status, the path change components are read. Received from the path change information written by the state transition within the current scrolling window, the slow hidden state transition operator saves the source identifier for each dimension of the read path continuation component when generating the macroscopic load path state after the contraction constraint. , Indicates the continuation of the path after reading. The source identifier for each dimension includes values ​​for upward continuation, downward continuation, stable continuation, and mismatch state. If the first dimension... The path continuation components after reading each dimension originate from the slow hidden state preserved components that carry the current macroscopic load continuation state, then... Record the current macroeconomic load continuation status. If the first If the path continuation component after reading each dimension does not originate from the slow hidden state retention component carrying the current macroscopic load continuation state, then... Records are in a mismatch state, based on all The generation path continues the source state of the components. , This indicates the current macroscopic load continuation status identifier corresponding to the read path continuation component, and sets a stability constant. , This represents a positive number with a zero denominator in the calculation of the determination quantity used to judge the effective amplitude of the path continuation component and avoid sudden changes in the load path. Its dimension is the hidden state numerical amplitude. It is valid when at least one condition is met. The path after reading continues along the component dimension, and all satisfy... The dimension corresponding to All are equal to the current macroeconomic load continuation state At that time, Record as When there is no satisfying The path after reading continues along the component dimension, or there exists at least one that satisfies... The dimension corresponding to When the state is mismatched, The record is in a mismatch state; Based on the read path continuation components and the path change components after reading , forming path state difference quantity , This represents the dimension-wise difference between the path continuation component and the path change component in the macroscopic load path state after contraction constraints, for each dimension. ,Will and Take the absolute value of the difference to get , Indicates the first The path state difference value on each slow hidden state dimension, the path state difference quantity. Each field in the data is generated by the direct values ​​of the path continuation component and the path change component after reading. Based on path state differences The deviation of the path change component from the path continuation component is calculated, and the degree of deviation is used as the load path abrupt change criterion. , This represents the decision quantity used for comparison with load path abrupt change conditions, and sets the effective amplitude threshold for the path. And cite the stability constant , This represents the preset threshold used to exclude dimensions where neither path continuation nor path change components have formed a valid amplitude; its dimension is the hidden state numerical amplitude. This represents a positive number used to avoid division where the denominator is zero; its dimension is the hidden state numerical amplitude; it is the load path abrupt change judgment quantity. Determine as follows: ; in, This indicates the load path abrupt change determination value. The dimension index representing the slow hidden state. The dimension representing the slow hidden state. Indicates indexing from the slow hidden state dimension. to Accumulate. This indicates an indicator function that takes the value when the condition within the parentheses is true. The value is taken as follows when the condition inside the parentheses is not true. , This indicates that the path continuation component after reading is in the first position. The values ​​can be taken in the dimension of each slow hidden state. This indicates that the path change component after reading is in the first position. The values ​​can be taken in the dimension of each slow hidden state. express The absolute value, express The absolute value, Indicates the effective amplitude threshold of the path. Indicates the first The path state difference value on each slow hidden state dimension. This indicates taking the larger value among the values ​​within the parentheses. This represents the portion where the magnitude of the path change component exceeds the magnitude of the path continuation component. Represents the stability constant, a constant and constant It is a dimensionless constant, in the formula , , , and All values ​​are latent state numerical amplitudes. The amplification term within square brackets in the formula is a dimensionless ratio. Both the numerator and denominator of the formula are cumulative latent state numerical amplitudes. (Load path abrupt change determination quantity) The value is dimensionless. Set load path mutation threshold , This represents the preset threshold in the load path abrupt change condition, with a value that is a dimensionless positive number, and is used to determine the load path abrupt change. With load path mutation threshold The comparison is used to determine the load path mutation. , This indicates the result of the load path abrupt change determination corresponding to the current scrolling window, when the path continues to the source status. Equal to the current macroeconomic load continuation state And load path mutation determination quantity Reaching the load path mutation threshold At that time, the result of the load path mutation determination will be... To satisfy the load path abrupt change condition, when the source state of the path continuation component. This does not equate to the current state of macroeconomic load continuity. Or load path mutation determination quantity The load path mutation threshold was not reached. At that time, the result of the load path mutation determination will be... Set as not meeting the load path mutation condition; In the result of load path mutation determination This indicates that, under the condition that the macroscopic load path state after contraction constraints meets the load path abrupt change condition, the current rolling window is read. The start sampling time and the time to terminate sampling The start and end sampling times are bound to the periods of abrupt changes in the macroscopic load continuation path. , This indicates the period of abrupt change in the macroscopic load continuity path, determined by the start and end sampling times of the current rolling window. If the load path abrupt change determination result... If the macroscopic load path state after the contraction constraint does not meet the load path abrupt change condition, then no macroscopic load path abrupt change period will be generated, and the load path abrupt change determination result will be retained. The unsatisfied state.

[0029] like Figure 7 As shown, step S7 specifically involves: Before matching the start and end sampling times, read the load path abrupt change determination results; When the load path mutation determination result is that the load path mutation condition is not met, the macro load continuation path mutation period is set to an empty period identifier, the path status output content is set to empty content, and a power load prediction result containing the load path mutation determination result and the macro load path status after contraction constraint is generated. When the load path mutation determination result is that the load path mutation condition is not met, the start and end sampling time matching, path state segment extraction and output range constraint are not executed. When the load path mutation determination result is that the load path mutation condition is met, the start and end sampling times of the macro load path state after the contraction constraint are matched according to the macro load continuation path mutation period. During periods of abrupt changes in the macro-load continuation path The macroscopic load path state after contraction constraints is the input object. This indicates the period of abrupt change in the macroscopic load continuation path, determined by the start and end sampling times of the current rolling window. This indicates the starting sampling time during the period of abrupt change in the macroscopic load continuation path. This indicates the termination sampling time during the period of abrupt change in the macro load continuation path; the current rolling window is represented as... The sampling time sequence within the current scrolling window is represented as follows: , This indicates the current length of the scrolling window. This indicates the index of the starting sampling time of the current scrolling window. Indicates index The corresponding sampling time; To match start and end sampling times, a state index mapping is established based on the sampling time index of the current rolling window to generate the macroscopic load path state after contraction constraints. , Indicates the sampling time The corresponding macroscopic load path state after contraction constraints, each Represented according to a fixed field order , dimension , The dimension representing the slow hidden state. Indicates the sampling time The corresponding path continuation component after reading, with dimension 1 , Indicates the sampling time The corresponding path changes after reading, with dimensions of When the slow hidden state transition operator generates the macroscopic load path state after window-level shrinkage constraint, the macroscopic load path state after window-level shrinkage constraint is compared with each sampling time in the current rolling window. Perform index binding to form This ensures that the start and end sampling times match objects with clearly defined states. Based on the abrupt change in the macro load continuation path Match the start and end sampling times of the sampling time sequence within the current scrolling window to satisfy... The sampling time index is preserved to form a set of indexes for abrupt change periods. , This represents the set of sampling time indices that match the abrupt changes in the macroeconomic load continuation path, arranged in ascending order of sampling time. and from Read each corresponding , forming path state fragments , This represents a state segment in the macro load path state after contraction constraints that matches the start and end sampling times of the abrupt change period in the macro load continuation path. Based on path status fragments The path continuation and path change components corresponding to each sampling time are read one by one to form the path state during the abrupt change period. , This indicates the path status during the abrupt change in the corresponding macroeconomic load continuity path. Each sampling time corresponds to one A dimensional state vector, the first dimensional state vector Each field represents a path continuation component, and the state vector is followed by... Each field represents a path variation component; Based on the path status during the mutation period The output range constraint is applied to the macroscopic load path state after the contraction constraint, and the output range constraint limits the sampling time index to... ,reserve The path continuation and path change components corresponding to each sampling time are included, excluding the current scrolling window. China does not belong to The path state corresponding to the sampling time is used to form the path state output content after being constrained by the output range. The path status output is sorted in ascending order by sampling time, and each item contains a sampling time field and a... Dimensional path continuation component field and one Dimensional path change component field; Based on the path status output and the macro load duration of path abrupt changes. This generates electricity load forecast results, which include periods of abrupt changes in the macroscopic load continuation path. Macroscopic load path state after contraction constraints The path status output includes the sampling time field, path continuation component field, and path change component field. The macro load continuation path mutation period in the power load forecast result is used to identify the sampling time range of macro load path mutations occurring in the target forecast area. The macro load path status after contraction constraint is used to retain the complete path status formed by the hidden state transition model. The path status output content is used to limit the path continuation component and path change component that can be read within the macro load continuation path mutation period.

[0030] like Figure 9 As shown in the figure, this diagram illustrates how micro-increment records, after undergoing direction determination, amplitude determination, and same-direction filtering, form same-direction increment objects. The shrinkage amount at the location of the micro-increment is then input into the state retention ratio generation process. The state retention ratio further influences the state transition process, causing the retention and writing of slow-hidden states to be constrained by the concentrated changes in micro-increments, ultimately forming the macro-load path state after shrinkage constraints.

[0031] like Figure 10 As shown in the figure, this diagram illustrates the extraction process of path continuity and path change components in the macroscopic load path state after contraction constraints. By calculating the differences between continuous and changed state segments, the path state difference is obtained. The path state difference is then compared with a threshold to form the load path abrupt change judgment result, and further outputs the corresponding time period state or empty time period result. like Figure 11 As shown in the figure, this figure illustrates how the total load sequence is truncated by a rolling window to form a macroscopic window sequence. The current macroscopic load continuity status is determined based on directional accumulation and threshold comparison. At the same time, it shows how multiple microscopic electricity consumption object load sequences are aligned by time and incrementally statistically analyzed to form microscopic electricity consumption object load increment records. This provides input for determining the dominant incremental direction of the window and generating the contraction amount at the location of microscopic increments.

[0032] Example 2: To verify the effectiveness of the proposed power load forecasting method based on micro-incremental contraction and hidden state transition model in identifying load path mutations in the target prediction area, experiments were set up to verify the following processes in the invention: current macro-load continuity state generation process, micro-electricity object load increment record generation process, load increment occurrence location determination process, same-direction increment object screening process, micro-increment occurrence location contraction amount generation process, state retention ratio generation process, slow hidden state transition operator, and load path mutation judgment process. This experiment focuses on verifying whether the shrinkage of the generated micro-increment location can be effectively embedded in the state retention ratio generation process of the slow hidden state transition operator when the location of the same-direction load increment of multiple micro-electricity objects changes from dispersed to concentrated, and whether it can constrain the degree of retention of the original macro load path tendency by the slow hidden state, so that the macro load path state after shrinkage constraint can more fully reflect the path change components under the appearance of macro load continuation. I. Experimental Tools and Operating Conditions The experiment was conducted in the same computing environment. The training set, validation set and test set all used the same sampling time alignment method, rolling window truncation method and preset threshold setting. During the experiment, the sampling period, window length, current rolling window and previous rolling window truncation rules and load path change conditions remained unchanged. Only the state coding layer parameters, incremental coding layer parameters, linear mapping parameters in the state retention ratio generation process and candidate writing linear mapping layer parameters were updated. Table 1 Experimental tools and operating conditions

[0033] II. Experimental Data and Division The experiment used the historical total load sequence of the target prediction area and the load sequences of 10 micro-electricity objects. The historical sampling time range was from 00:00 on May 1, 2024 to 23:45 on June 30, 2024, with a uniform sampling period of 15 minutes, resulting in a total of 5856 historical sampling points. The micro-electricity object indices were U1, U2, U3, U4, U5, U6, U7, U8, U9, and U10. The historical total load sequence included the sampling time and total load value, while the micro-electricity object load sequence included the micro-electricity object index, sampling time, and load value. The historical total load series ranges from 81.6MW to 174.5MW, with an average load of 126.8MW and a standard deviation of 18.4MW. The average load of the 10 micro-level electricity consumers ranges from 4.1MW to 14.2MW, with U4 having a higher average load of 14.2MW and U7 having a lower average load of 4.1MW. The load series of multiple micro-level electricity consumers are aligned with the sampling time of the historical total load series, so that each historical rolling window can simultaneously obtain the total load value of the target prediction area and the load values ​​of multiple micro-level electricity consumers. The current scrolling window has a length of 12 sampling points, the previous scrolling window has a length of 12 sampling points, and the current scrolling window has a sliding step of 4 sampling points. Each historical scrolling window includes a previous scrolling window and a current scrolling window. The training set, validation set, and test set are divided consecutively according to the order of sampling time, and there is no overlap in sampling time. The same historical scrolling window does not cross the sampling time range of the training set, validation set, and test set. Table 2. Division of experimental data and distribution of labeled states

[0034] In the test set, taking test sample-0094 as an example, its previous rolling window was from 06:00 on June 25, 2024 to 08:45 on June 25, 2024, and the current rolling window is from 09:00 on June 25, 2024 to 11:45 on June 25, 2024. Within this historical rolling window, the total load value increased from 128.7MW to 154.0MW. Within the current rolling window, U1 increased from 12.9MW to 15.9MW, and U2 increased from 8... U4 increased from 14.8MW to 17.9MW, U6 increased from 10.1MW to 12.6MW, and U9 increased from 10.5MW to 13.1MW. The current macro load continuation status corresponding to this window is an upward continuation. The shrinkage of the location of the micro increment is 5.84 sampling intervals. The macro load path status has been marked as an upward continuation corresponding to the path continuation component, and the path change component has been written. The load path abrupt change condition is met. like Figure 8 As shown in the figure, this figure illustrates the locations where load increments occur for multiple micro-level electricity users within the previous and current scrolling windows. The blue and orange positions correspond to the increment location markers for different micro-level electricity users. By comparing the position contraction reference span within the previous scrolling window with the reduction in the position span within the current scrolling window, the contraction amount of the micro-level increment location is formed, which is used to characterize the process of the same-direction increment objects changing from dispersed increments to concentrated increments.

[0035] like Figure 12 As shown, this figure illustrates the load change process of ten micro-level electricity users from 2:00 PM to 4:45 PM on June 25, 2024. Each sub-figure corresponds to one micro-level electricity user. The vertical markers are used to indicate the location of the identified increment within the current window, reflecting the supporting role of local changes in micro-level objects in determining path abrupt changes.

[0036] III. Initialization of Experimental and Model Parameters In the experiment, the generation of the current macro load continuity state adopts the total load amplitude threshold, cumulative ratio threshold, and consecutive number threshold. The generation of the load increment record of micro-electricity objects adopts the micro load amplitude threshold. The determination of the load increment location adopts the candidate increment amplitude threshold, the number of direction-consistent objects threshold, and the proportion of direction-consistent objects threshold. The screening of same-direction increment objects adopts the same-direction ratio threshold. The generation of the shrinkage amount at the micro increment location adopts the shrinkage judgment threshold. When the shrinkage amount at the micro increment location enters the state retention ratio generation process, the shrinkage amount threshold, the preset ratio lower limit, the preset ratio upper limit, and the retention range threshold are adopted. The hidden state transition model includes a state encoding layer, an incremental encoding layer, a slow hidden state, a fast hidden state, a slow hidden state transition operator, and a candidate write linear mapping layer. The slow hidden state transition operator internally includes a state retention ratio generation process. The state encoding layer is used to convert the current macro load continuation state into a slow hidden state. The incremental encoding layer is used to convert the micro load increment record of the electricity consumption object into a fast hidden state. The state retention ratio generation process is used to generate a basic state retention ratio and form a shrinkage constraint quantity for the basic state retention ratio based on the shrinkage amount at the location of the micro increment. The slow hidden state transition operator generates a state transition control quantity based on the state retention ratio after embedding the shrinkage amount at the location of the micro increment, and further generates the macro load path state after shrinkage constraint. Table 3 Initialization settings for experimental and model parameters

[0037] IV. Training process and loss iteration records The training samples are generated from the historical rolling window in the training set. Each training sample includes the current macro load continuation status, the micro load increment record of the electricity user, and the shrinkage amount at the location of the micro increment. Each training sample corresponds to a labeled macro load path status. The labeled macro load path status is saved according to the field organization method of path continuation component and path change component, and is used as the comparison target of training differences. After the training samples are input into the hidden state transition model, the current macroscopic load continuation state enters the state encoding layer to form a slow hidden state, and the microscopic electricity object load increment record enters the increment encoding layer to form a fast hidden state. The slow hidden state and the fast hidden state enter the state retention ratio generation process to form the basic state retention ratio. The shrinkage amount at the location where the microscopic increment occurs enters the state retention ratio generation process to form a shrinkage constraint amount. Based on the shrinkage constraint amount, the state retention ratio after embedding the shrinkage amount at the location where the microscopic increment occurs is formed. Subsequently, the slow hidden state transition operator generates a state transition control quantity based on the state retention ratio after the shrinkage of the embedded micro-increment location, and performs state transition on the slow hidden state according to the state transition control quantity to generate a slow hidden state after shrinkage constraint. The path continuation component and path change component are extracted from the slow hidden state after shrinkage constraint to form a macro load path candidate state, and the macro load path state after shrinkage constraint is further generated. The macro load path state after shrinkage constraint is compared with the labeled macro load path state to form a training difference. The training difference is used as a loss record in the training process to update the state encoding layer parameters, the incremental encoding layer parameters, the linear mapping parameters in the state retention ratio generation process, and the candidate writing linear mapping layer parameters. After each training round, the validation samples in the validation set are input into the current hidden state transition model. During the validation set processing, the model parameters are not updated based on the validation set differences. The validation set differences are only used to determine whether the training stop condition is met. When the preset number of training rounds is reached, or the validation set differences do not decrease within the preset number of consecutive rounds, training is stopped and the corresponding model parameters are fixed. Table 4. Iteration Record of Training Differences and Validation Set Differences for the Complete Scheme

[0038] As shown in Table 4, the training difference gradually decreases with the increase of training rounds. The validation set difference reaches its lowest state in the 108th round. After the 108th round, the validation set difference does not continue to decrease and does not form a new decreasing state within the preset number of consecutive rounds. Therefore, the model parameters corresponding to the 108th round are fixed as the model parameters of the complete scheme. V. Comparison Scheme Setup and Test Set Results To verify the role of the generation process of the shrinkage amount at the location of the micro-increment and the retention ratio of the embedded state, four sets of comparison schemes were set up. Compared with Scheme 1, which only forms the current macro load continuity state based on the total load sequence and performs state transition based on the slow hidden state, it does not use the load increment record of micro-electricity objects, does not generate the shrinkage amount of the micro increment occurrence location, and does not embed the shrinkage amount of the micro increment occurrence location into the state retention ratio generation process. Compared with Scheme 2, which records the load increment of micro-electricity objects as the input of fast-concealing state, but does not generate the shrinkage amount of the location where the micro-increment occurs, the fast-concealing state carries the local load change information within the current rolling window, but does not form the constraint of the shrinkage of the location where the load increment occurs across the window. Compared with Scheme 3, which generates the shrinkage amount at the location of the micro-increment but does not embed it into the state retention ratio generation process, the shrinkage amount at the location of the micro-increment does not directly constrain the degree to which the slow hidden state retains the original macro-load path tendency. All four schemes use the same training set, validation set, and test set partitioning, the same current scroll window length, previous scroll window length, number of training rounds, and training stop condition. Except for the differences mentioned above, all other parameters remain the same. The test set contains a total of 235 historical rolling windows, of which 39 historical rolling windows have been marked as meeting the load path change conditions and 196 historical rolling windows have been marked as not meeting the load path change conditions. During the test set processing, the model parameters are not updated, the preset thresholds are not adjusted, and the load path change conditions are not changed. like Figure 13As shown in the figure, this figure illustrates the continuous changes in the contraction amount at the location of the micro-increment and the load path mutation index within 235 scrolling windows of the test set. The upper line corresponds to the contraction amount at the location of the micro-increment, and the lower line corresponds to the load path mutation index. The figure marks the window numbers and corresponding values ​​that meet the mutation conditions, which are used to reflect the experimental results that the path mutation index increases synchronously when the contraction amount increases.

[0039] Table 5 Comparison results of test sets for different schemes

[0040] As can be seen from Table 5, compared with Scheme 1, which only uses the total load sequence to form the current macro load continuation state, it is difficult to identify path abrupt changes caused by the synchronous incremental concentrated changes of multiple micro-electricity objects when the macro load still shows a continuation state. Although Scheme 2 introduces the load increment records of micro-electricity objects to form a fast concealed state, it does not generate the shrinkage amount of the micro-increment occurrence location. Therefore, the local load change information in the fast concealed state cannot directly constrain the state retention degree of the slow concealed state. Although Scheme 3 generates the shrinkage amount of the micro-increment occurrence location, it does not embed it into the state retention ratio generation process, which causes the shrinkage amount of the micro-increment occurrence location to not directly affect the state retention ratio of the slow concealed state. The complete scheme embeds the shrinkage amount of the micro-increment occurrence location into the state retention ratio generation process, which can reduce the inertial retention of the original macro load path tendency of the slow concealed state and increase the writing ratio of path change components. Therefore, the average path state difference is lower, and the abrupt change judgment accuracy, recall rate and F1 value are higher. VI. Ablation Experiment To further verify the role of each component of the present invention in the final power load prediction result, an ablation experiment was set up based on the complete scheme. The training set, validation set and test set division remained unchanged in each ablation setting, and the current rolling window length, the previous rolling window length, the unified sampling period, the initial value range of model parameters and the training process control data remained unchanged. Only the corresponding processing links were removed or changed. Ablation 1 does not filter objects with the same direction of increment based on the window's dominant increment direction, which is used to verify the role of the window's dominant increment direction in the determination process of objects with the same direction of increment. Ablation 2 does not use the location of the load increment occurrence within the previous rolling window, and is used to verify the effect of the contraction amount of the load increment occurrence location across the window; The ablation step 3 sets the shrinkage amount at the location where the micro-increase occurs to zero to verify the overall effect of the shrinkage amount at the location where the micro-increase occurs; The ablation process retains the micro-incremental location shrinkage amount, but does not embed it into the state retention ratio generation process, which is used to verify the role of the embedded state retention ratio generation process. The five non-use fast-concealing states are used to verify the role of the load increment record of micro-electricity objects in forming fast-concealing states; Ablation 6 does not add the slow hidden state write component according to the state transition control during state transition, which is used to verify the effect of the slow hidden state write component on the formation of path change components. Table 6 Ablation Experiment Results

[0041] As shown in Table 6, the complete scheme outperforms all ablation settings in terms of validation set difference, average path state difference, mutation detection accuracy, recall, and F1 score. The ablation method removes the window-dominated incremental direction from the screening of same-direction incremental objects, allowing some micro-electricity objects that are inconsistent with the current macro load continuation state to also participate in the location contraction judgment. This leads to an increase in the number of cases where the marked conditions are not met but the output conditions are met, indicating that the window-dominated incremental direction is necessary to limit same-direction incremental objects. Ablation 2 does not use the load increment occurrence location within the previous rolling window, so the shrinkage of the micro increment occurrence location cannot reflect the change process from dispersion to concentration across the window, resulting in a decrease in recall and F1 value. This indicates that cross-window load increment occurrence location matching is an important basis for forming the shrinkage of the micro increment occurrence location. When ablation 3 sets the shrinkage amount at the location of the micro-increment to zero, the slow hidden state transition operator can only perform state transitions based on the retention ratio of the basic state and cannot receive the constraint of the shrinkage amount at the location of the micro-increment on the retention degree of the slow hidden state, resulting in insufficient writing of path change components. This indicates that the shrinkage amount at the location of the micro-increment has a direct effect on the determination of the abrupt change period of the macro load continuation path. Although ablation four retains the generation of the contraction amount at the location of the micro-increment, it does not embed it into the state retention ratio generation process. This means that the contraction amount at the location of the micro-increment cannot directly reduce the degree of retention of the original macro-load path tendency of the slow hidden state, resulting in the macro-load path state after contraction constraint not being able to fully absorb the path change information within the current rolling window. The ablation of the fast hidden state prevents the load increment records of micro-level electricity users from being used as local load change information within the current rolling window to enter the slow hidden state transfer operator, resulting in an increase in the average path state difference and a decrease in the accuracy and recall of mutation judgment. This indicates that the fast hidden state is an important state input for carrying micro-level load change information. Ablation 6 does not increase the slow hidden state write component according to the state transition control quantity during state transition, which makes it difficult for the path change components introduced by the fast hidden state in the candidate slow hidden state to be fully written into the slow hidden state. This increases the number of cases where the labeled conditions are met but the output conditions are not met, indicating that the slow hidden state write component is necessary for forming readable path change components. VII. Comparison Results and Conclusions The above comparative and ablation experiments show that the shrinkage amount at the location of the micro-increment in this invention is not used as a common input field to participate in the model processing, but is embedded in the state retention ratio generation process of the slow hidden state transition operator, which is used to directly constrain the degree of retention of the slow hidden state on the original macro load path tendency. Compared to the comparison schemes that only use the total load sequence, the comparison schemes that only use the load increment records of micro-level electricity users, and the comparison schemes that do not embed the shrinkage amount of the location of the micro-level increment into the state retention ratio generation process, the complete scheme can generate a lower average path state difference and improve the consistency of the determination of the abrupt change period of the macro-level load continuation path. Experimental results further illustrate that the key steps in the load path mutation identification process of this invention are: same-direction incremental object screening, cross-window load increment occurrence location matching, micro-increment occurrence location contraction generation, state retention ratio embedding, fast concealed state input, and slow concealed state writing components. After the above steps work together, when the total load in the target prediction area still shows an upward continuation, a downward continuation, or a stable continuation, the change in the location of multiple micro-electricity objects' load increment occurrence from dispersed to concentrated is transformed into a constraint on the degree of slow concealed state retention. This allows the macro-load path state after contraction constraint to more fully accept the path change information within the current rolling window, thereby generating a macro-load continuation path mutation period and outputting it as the power load prediction result.

[0042] This invention addresses the difficulty in identifying path abrupt changes caused by the concentrated synchronous increments of micro-level electricity users when the overall macro load continues to show a sustained trend in existing power load forecasting scenarios. By screening for unidirectional incremental electricity users and calculating the contraction amount at the location of their load increments across the rolling window, this contraction amount is used as a constraint to embed into the state retention ratio generation process of the slow hidden state transition operator in the hidden state transition model. The inertial retention degree of the macro load path state and the writing ratio of local load change information are dynamically adjusted according to the degree of micro-level increments shifting from dispersed to concentrated. This allows for the determination of macro load path abrupt change periods even under the appearance of a sustained overall load, outputting load forecast results that include the abrupt change period and the path state after contraction constraints. This helps improve the identification effect of load path abrupt changes. It solves the problem that existing methods struggle to effectively identify macro load path abrupt change periods based on the shift from dispersed to concentrated unidirectional load increments of micro-level electricity users when the overall macro load continues to show a sustained trend in power load forecasting scenarios.

[0043] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A power load forecasting method based on a micro-incremental contraction and hidden state transition model, characterized in that, Includes the following steps: S1. Obtain the total load sequence and multiple micro-level electricity consumption object load sequences of the target prediction area, form the current macro-level load continuity status based on the total load sequence, and form the micro-level electricity consumption object load increment record based on the multiple micro-level electricity consumption object load sequences. S2. Based on the load increment records of micro-level electricity users, determine the location of the load increment of each micro-level electricity user within the current rolling window, and determine the dominant increment direction of the window based on the current macro-level load continuity status. S3. Based on the location of the load increment and the dominant increment direction of the window, determine the same-direction increment object, and generate the micro increment location contraction amount based on the load increment location of the same-direction increment object in the current scrolling window and the previous scrolling window. S4. Construct a hidden state transition model, taking the current macroscopic load continuation state as the slow hidden state, taking the microscopic electricity object load increment record as the fast hidden state, and embedding the shrinkage amount of the microscopic increment location into the state retention ratio generation process of the slow hidden state transition operator. S5. Based on the state retention ratio after the shrinkage of the embedded micro-increment, the slow hidden state is transferred to generate the macro load path state after the shrinkage constraint. S6. Determine the load path abrupt change based on the macro load path state after the contraction constraint. When the macro load path state after the contraction constraint meets the load path abrupt change condition, generate the macro load continuation path abrupt change period. S7. Outputs power load forecast results including the period of abrupt change in the macro load continuation path and the state of the macro load path after contraction constraints.

2. The power load forecasting method based on the micro-incremental contraction and hidden state transition model according to claim 1, characterized in that, Step S1 specifically includes: The total load sequence is sampled according to the current rolling window and the previous rolling window to form a macroscopic window sequence; the load sequences of multiple microscopic electricity users are time-aligned according to the sampling time of the macroscopic window sequence to form a microscopic electricity user aligned load sequence. Calculate the direction of total load change between adjacent sampling times within the current rolling window and generate a cumulative result. Compare the cumulative result with a preset threshold to determine the current macroscopic load continuation status. Based on the aligned load sequence of micro-electricity users, the direction and magnitude of load increment for each micro-electricity user between adjacent sampling times are calculated to form a load increment record for each micro-electricity user.

3. The power load forecasting method based on the micro-incremental contraction and hidden state transition model according to claim 1, characterized in that, Step S2 specifically includes: Extract the load increment direction and load increment magnitude of each micro-electricity object at each sampling moment within the current rolling window from the load increment record of each micro-electricity object, and form the current window increment time sequence; The load increment magnitude is compared with the preset candidate increment magnitude threshold to obtain a set of candidate increment locations for each micro-electricity consumer; the load increment direction within the candidate increment location set is screened for consistency to generate the load increment location of each micro-electricity consumer within the current scrolling window; The load path continuation direction is obtained by matching the current macro load continuation status, and this direction is determined as the window-dominant incremental direction.

4. The power load forecasting method based on the micro-incremental contraction and hidden state transition model according to claim 1, characterized in that, Step S3 specifically includes: Extract the load increment direction corresponding to the location where the load increment occurs for each micro-level electricity consumer and form a location-direction correspondence; match the load increment direction of each micro-level electricity consumer with the dominant increment direction of the window, calculate the proportion of consistent directions, and filter to obtain the same-direction increment objects and the same-direction location set; Get the position of the same direction load increment of the same object in the previous scroll window, and perform sequence matching between the position of the same direction load increment in the current scroll window and the previous scroll window to form cross-window same direction position pairs. Based on the cross-window same-direction position pairs, determine the position contraction reference span of the previous scrolling window and the current window position span of the current scrolling window respectively, and compare them to obtain the position span reduction amount; compare the position span reduction amount with the preset contraction judgment threshold, and when the condition is met, generate a micro increment based on the position span reduction amount to generate the position contraction amount.

5. The power load forecasting method based on the micro-incremental contraction and hidden state transition model according to claim 1, characterized in that, Step S4 specifically includes: Construct a hidden state transition model that includes slow hidden state, fast hidden state, and slow hidden state transition operator, and limit the calculation process in the slow hidden state transition operator used to determine the degree of preservation of slow hidden state to the state preservation ratio generation process. The current macro load continuity state is encoded to form a slow-concealed state that carries the macro load path tendency; the load increment records of micro-electricity objects are incrementally encoded to form a fast-concealed state that carries the local load change information of the current rolling window. The basic state retention ratio is calculated based on the slow hidden state and the fast hidden state; the shrinkage amount at the location where the micro increment occurs is input into the state retention ratio generation process to generate a shrinkage constraint amount for the basic state retention ratio; the basic state retention ratio is updated based on the shrinkage constraint amount to obtain the state retention ratio after embedding the shrinkage amount at the location where the micro increment occurs and is written into the slow hidden state transition operator.

6. The power load forecasting method based on the micro-incremental contraction and hidden state transition model according to claim 1, characterized in that, Step S5 specifically includes: The state transition control quantity is generated based on the state retention ratio after the shrinkage of the embedded micro-increment location; the path continuation component in the slow hidden state is retained based on the state transition control quantity to form the slow hidden state retention component. Combine the fast hidden state and the slow hidden state to perform candidate write calculations to form a candidate slow hidden state; determine the state write ratio according to the state transition control quantity, and perform write calculations on the path change components in the candidate slow hidden state to form the slow hidden state write component. By combining the slow hidden state retention component and the slow hidden state write component, a slow hidden state after contraction constraint is generated. The path continuation component and path change component are extracted to form a macro load path candidate state. After reading the path state, a macro load path state after contraction constraint is generated.

7. The power load forecasting method based on the micro-incremental contraction and hidden state transition model according to claim 1, characterized in that, Step S6 specifically includes: Extract the path continuation component and the path change component from the macroscopic load path state after contraction constraints, and calculate the path state difference. The deviation of the path change component from the path continuation component is calculated based on the path state difference, forming the load path abrupt change judgment quantity; The load path mutation judgment quantity is compared with the preset threshold in the load path mutation condition. When the mutation condition is met, the macro load continuity path mutation period is generated according to the start and end sampling times of the current rolling window.

8. The power load forecasting method based on the micro-incremental contraction and hidden state transition model according to claim 1, characterized in that, Step S7 specifically includes: Based on the start and end sampling times of the abrupt change period of the macro load continuation path, the corresponding path state segments are matched and extracted from the macro load path state after the contraction constraint to form the path state of the abrupt change period. The output range is constrained based on the path state during the abrupt change period, resulting in path state output content that includes both path continuation and path change components within the abrupt change period. By combining the path status output with the period of abrupt changes in the macro load continuation path, the power load forecast results are generated and output.

9. The power load forecasting method based on the micro-incremental contraction and hidden state transition model according to claim 4, characterized in that, Before forming cross-window same-direction position pairs, it is first determined whether the same-direction incremental objects have load increment occurrence positions consistent with the dominant increment direction of the window in both the current scrolling window and the previous scrolling window. Only same-direction incremental objects that have same-direction load increment occurrence positions in both windows are used to form cross-window same-direction position pairs. The load increment occurrence positions participating in the calculation of position shrinkage reference span, position span reduction, and micro increment occurrence position shrinkage are all limited to load increment occurrence positions consistent with the dominant increment direction of the window.

10. The power load forecasting method based on the micro-incremental contraction and hidden state transition model according to claim 5, characterized in that, When embedding the shrinkage amount at the location of the micro-increment into the state retention ratio generation process, the slow hidden state and the fast hidden state are kept as the state inputs of the slow hidden state transition operator. The shrinkage amount at the location of the micro-increment is only used to generate the shrinkage constraint amount to update the basic state retention ratio. The degree of retention of the slow hidden state is constrained by the state retention ratio generation process.

Images