A dynamic power load interval prediction method and system based on long-short term windows

Abstract

This invention relates to a dynamic power load interval prediction method and system based on long and short time windows, belonging to the field of power load interval prediction. The method includes: acquiring a historical load dataset and dividing it into a training set and a calibration set; inputting the calibration set into a pre-trained MoE model to output load prediction values, determining the upper and lower bounds of the load interval, and calculating the prediction boundary residual sequence; constructing long and short time windows and setting long and short forgetting factors, and calculating long and short boundary weights; obtaining long and short critical prediction boundary residuals based on the target fault tolerance rate and long and short boundary weights; extracting long and short prediction boundary residuals based on the index of the long and short critical prediction boundary residuals, and calculating the global safety margin; obtaining the power load interval based on the global safety margin and the upper and lower bounds of the load interval. This invention can capture the real fluctuation pattern of load data, adaptively adjust the width of the prediction interval, and realize dynamic load interval prediction.

Description

Technical Field

[0001] This invention relates to the field of power load interval forecasting, and mainly to a dynamic power load interval forecasting method and system based on long and short windows. Background Technology

[0002] Power load forecasting is a crucial support for ensuring the safe dispatch, economical operation, and risk prevention of the power system. Traditional load forecasting is mostly based on point forecasting, outputting only a single load forecast value, which cannot reflect the uncertainties brought about by load fluctuations, sudden weather changes, and changes in electricity consumption behavior, and is difficult to meet the actual operation needs of the power grid. In contrast, power load interval forecasting can provide the range of load fluctuations at a specified confidence level, making up for the shortcomings of point forecasting in expressing uncertainty. Therefore, to meet the security requirements of power forecasting, outputting load forecast intervals with upper and lower bounds is essential.

[0003] Chinese invention patent application CN113139671A discloses a method and system for predicting power load intervals based on a branch structure model. This technical solution includes dividing pre-collected historical power load data into different datasets according to categories. Each dataset is input into parallel GRU modules in a pre-established and trained branch structure model, which outputs power load interval prediction results. The branch structure model includes a number of parallel GRU modules corresponding to the number of categories, and the outputs of each GRU module are connected to a fully connected layer. The daily scheduling plan for the next day is arranged using the interval range of the power load interval prediction results. However, the above technical solution only directly outputs the power load interval prediction through parallel GRU modules and a fully connected layer. As a result, without any residual-based dynamic calibration, weight update, and interval correction mechanisms, the output interval remains fixed once generated and cannot be adaptively adjusted according to real-time prediction errors. Furthermore, the GRU modules in the aforementioned technical solution operate in fixed parallelism, failing to adaptively allocate computing resources based on input features. This easily leads to some modules being overloaded while others remain idle, resulting in poor model efficiency and stability. Additionally, the aforementioned technical solution cannot dynamically generate a global safety margin based on the target fault tolerance rate, and cannot adaptively tighten or loosen the interval while ensuring a specified coverage rate. This easily leads to problems such as insufficient coverage due to overly narrow intervals or scheduling redundancy due to overly wide intervals.

[0004] Therefore, there is an urgent need for a dynamic power load range prediction method that can achieve dynamic calibration of the prediction range. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a dynamic power load interval prediction method based on long and short window periods.

[0006] The technical solution of the present invention is as follows: On the one hand, this invention proposes a dynamic power load interval prediction method based on long and short-term windows, the method comprising: Acquire historical load datasets of target monitoring nodes in the power system and preprocess them. Divide the preprocessed historical load datasets into training sets and calibration sets according to a preset ratio. The calibration set is input into the hybrid expert MoE model pre-trained with the training set to predict the load and output the predicted load value. The upper and lower bounds of the load interval are determined based on the predicted load value, and the prediction boundary residual sequence is calculated using the upper and lower bounds of the load interval. Construct long and short time windows and set corresponding long and short forgetting factors. Calculate the corresponding long and short boundary weights based on the long and short forgetting factors. Obtain the long and short critical prediction boundary residuals based on the preset target tolerance rate and long and short boundary weights. Based on the index of the long- and short-term critical prediction boundary residuals, the long- and short-term prediction boundary residuals are extracted from the prediction boundary residual sequence. The global safety margin is calculated using the long- and short-term prediction boundary residuals. The power load range is obtained based on the global safety margin and the upper and lower boundaries of the load range.

[0007] Preferably, the historical load dataset includes historical load observation sequences, multi-site environmental meteorological covariates, and time attribute labels; The preprocessing includes label mapping, feature construction, and long-tailed distribution sample augmentation, wherein: Specifically, the label mapping uses a trigonometric function encoding mechanism to map time attribute labels to historical load observation sequences, in order to preserve the periodicity and continuity of historical load observation sequences. The feature construction specifically involves extracting load statistics parameters based on a preset sliding window mechanism, and then concatenating them with the acquired historical lagged load observation sequences and historical multi-site environmental meteorological covariates to construct a time series feature vector. The long-tailed distribution sample augmentation of time series feature vectors based on generative adversarial networks and differential interpolation is performed as follows: Generative Adversarial Networks (GANs) are constructed based on time series data. The time series feature vectors are input into the GAN, and adversarial samples are generated based on the adversarial game process between the generator and the discriminator of the GAN. The adversarial examples are completed using a differential interpolation algorithm to generate a preprocessed historical payload dataset.

[0008] Preferably, the MoE model includes an expert layer, a gated routing layer, and a prediction layer, specifically: The expert layer includes multiple convolutional attention sub-networks based on contrastive learning as expert networks. Each expert network includes a one-dimensional convolutional layer and a Transformer layer containing a multi-head self-attention mechanism. Specifically, the one-dimensional convolutional layer is used to extract mutation features from the preprocessed historical load dataset, the contextual dependency of mutation features is established based on the Transformer layer containing a multi-head self-attention mechanism, and the initial prediction value corresponding to each expert network is output. The gated routing layer is constructed based on a multilayer perceptron. It extracts multi-site meteorological covariates from the preprocessed historical load dataset and fuses them with the acquired static features to obtain the gated input. The routing probability distribution of the gated input on each expert network is calculated based on the Softmax function. The prediction layer includes two quantile prediction heads and a Topology execution layer. The k-sparse activation mechanism calculates load forecasts based on the initial forecasts and routing probability distributions of each expert network, including upper quantile load forecasts and lower quantile load forecasts. The upper quantile load forecast is used as the upper limit of the load interval, and the lower quantile load forecast is used as the lower limit of the load interval.

[0009] Preferably, the prediction boundary residual sequence is calculated using the upper and lower bounds of the load interval, specifically as follows: Obtain the current load data and actual load observations; Calculate the difference between the lower bound of the load interval corresponding to the load data at the current moment and the actual load observation value, and use it as the first candidate prediction boundary residual; Calculate the difference between the actual load observation value and the upper bound of the load interval corresponding to the load data at the current time, and use it as the second candidate prediction boundary residual; The maximum value between the first candidate prediction boundary residual and the second candidate prediction boundary residual is selected as the prediction boundary residual at the current time. Sort the predicted boundary residuals at all times in ascending order to obtain the predicted boundary residual sequence.

[0010] Preferably, the construction of long and short time windows specifically involves setting the size to [value missing]. Short-term sliding window, size is The long-term sliding window and the corresponding long-term and short-term forgetting factors; The long-short boundary weights are obtained by calculating the ratio of the exponential decay term of the long-short forgetting factor at the current moment to the sum of the exponential decay terms at all moments within the preset time period.

[0011] Preferably, the short- and long-term critical prediction boundary residuals are obtained based on the preset target fault tolerance rate and short- and long-term boundary weights, specifically as follows: The long and short-term boundary weights corresponding to the predicted boundary residual sequence are accumulated until the accumulated weight first reaches or exceeds the target coverage, at which point the accumulation stops. The target coverage is the difference between 1 and the target tolerance. Extract the index of the last accumulated short-term boundary weight. ; Extract the index of the last accumulated long-term boundary weight. ; The first The prediction boundary residuals corresponding to each boundary weight are used as the short-term prediction boundary residuals. The first The prediction boundary residuals corresponding to each boundary weight are used as the long-term prediction boundary residuals. The global safety margin is obtained by calculating the short-term and long-term prediction boundary residuals based on adaptive fusion weighting.

[0012] Preferably, the adaptive fusion weight is calculated based on a preset sensitivity coefficient, the load fluctuation intensity index at the current moment, and a preset fluctuation threshold.

[0013] On the other hand, the present invention also provides a dynamic power load interval prediction system based on long and short windows, the system comprising: The data acquisition module acquires historical load datasets of target monitoring nodes in the power system and preprocesses them, dividing the preprocessed historical load datasets into training and calibration sets according to a preset ratio. The load boundary prediction module takes the calibration set as input to the hybrid expert MoE model pre-trained with the training set to perform load prediction and outputs the load prediction value; it determines the upper and lower bounds of the load interval based on the load prediction value and uses the upper and lower bounds of the load interval to calculate the prediction boundary residual sequence. The long-short-term prediction boundary residual calculation module constructs long-short-term time windows and sets corresponding long-short-term forgetting factors, calculates the corresponding long-short-term boundary weights based on the long-short-term forgetting factors, and obtains the long-short-term critical prediction boundary residuals based on the preset target fault tolerance rate and long-short-term boundary weights. Extract long and short-term prediction boundary residuals from the prediction boundary residual sequence based on the index of long and short-term critical prediction boundary residuals. The power load interval calculation module uses the long-term and short-term forecast boundary residuals to calculate the global safety margin, and obtains the power load interval based on the global safety margin and the upper and lower boundaries of the load interval.

[0014] In another aspect, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in the present invention.

[0015] In another aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method described in the present invention.

[0016] The present invention has the following beneficial effects: 1. This invention adopts a hybrid expert MoE model structure, using multiple parallel convolutional attention networks as expert layers, and combining gated routing and sparse activation mechanisms to adaptively extract features and distinguish patterns for different load modes, fully capturing the local mutation features and long-term temporal dependencies of load sequences; improving the feature representation capability and model generalization in complex load scenarios, and effectively improving the initial prediction accuracy of load intervals; 2. This invention uses parallel tracking with both long and short scale windows. It can maintain the statistical stability of the prediction interval by using the long-term window to avoid frequent false alarms caused by local minor disturbances. At the same time, it can use the short-term window in conjunction with the fluctuation intensity sensing mechanism to quickly widen the global safety margin in the early stages of extreme weather shocks or sudden changes in power consumption patterns, so as to realize the real-time adaptive adjustment of the prediction interval. This mechanism realizes the automation and high reliability of power load judgment standards across industries and scenarios through closed-loop feedback of error flow. 3. Based on the principle of target fault tolerance rate and critical residual screening, this invention dynamically calculates the global safety margin and corrects the upper and lower boundaries of the interval. While meeting the preset coverage requirements, it adaptively optimizes the interval width, ensuring a balance between interval reliability and compactness, improving the coverage and accuracy of load interval prediction, avoiding safety risks caused by excessively narrow intervals and waste of scheduling resources due to excessively wide intervals, and is more suitable for the real-time scheduling and safe operation requirements of the power grid. Attached Figure Description

[0017] Figure 1 This is a detailed flowchart of an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] It should be understood that the step numbers used in the text are for ease of description only and are not intended to limit the order in which the steps are performed.

[0020] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0021] The terms “comprising” and “including” indicate the presence of the described feature, whole, step, operation, element and / or component, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or collections thereof.

[0022] The term “and / or” refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes these combinations.

[0023] Example 1: See Figure 1 This invention provides a dynamic power load interval prediction method based on long and short window, the method comprising: S1. Obtain the historical load dataset of the target monitoring nodes of the power system and preprocess it. Divide the preprocessed historical load dataset into a training set and a calibration set according to a preset ratio. S11. The historical load dataset includes historical load observation sequences, multi-site environmental meteorological covariates, and time attribute labels. S12. The label mapping specifically employs a trigonometric function encoding mechanism to map time attribute labels to historical load observation sequences, thereby preserving the periodicity and continuity of the historical load observation sequences, wherein: S121. A trigonometric function encoding mechanism is used to map time attribute labels to historical load observation sequences. The calculation method is as follows: ; ; In the formula, A continuous sinusoidal mapping characteristic representing time attributes; A continuous cosine mapping feature representing time attributes; Indicates time; Indicates the preset target cycle length; Time attribute labels (such as 11 PM and 12 AM) differ greatly in terms of load values, but they are adjacent in terms of physical time. Mapping time using sine and cosine trigonometric functions can connect the beginning and end of time into a circle. Its main function is to preserve the periodicity and continuity of time characteristics, helping the model to better learn the daily / weekly cycle of load. S122. The feature construction specifically involves extracting load statistics parameters based on a preset sliding window mechanism, and then concatenating them with the acquired historical lagging load observation sequences and historical multi-site environmental meteorological covariates to construct a time series feature vector. The load statistical parameters refer to a set of dynamically derived features obtained by statistically calculating historical load sequences through a sliding window with a preset time step, used to quantify the smoothness and fluctuation intensity of power grid load changes; specifically, the load statistical parameters include, for example, the load mean used to characterize the basic physical inertia and recent smooth evolution trend of power grid load, and the load standard deviation used to quantify the fluctuation intensity and instability of recent load characteristics. The time series feature vector, for example, at a certain short-term load forecast time, specifically includes the following high-dimensional data sample: the mapped historical load observation sequence (e.g., the sine and cosine values ​​corresponding to 14:00), the current multi-site environmental meteorological covariates (e.g., temperature 32℃, humidity 80%, wind speed 3m / s), historical lag characteristics (e.g., the actual load of 500MW in the past hour, the actual load of 480MW at the same time yesterday), and rolling statistical parameters reflecting physical inertia (e.g., the load moving average of 490MW and the maximum peak-valley difference in the past 4 hours). These multidimensional data are concatenated together to form the time series feature vector input to the MoE model. S123. The long-tailed distribution samples are augmented to the time series feature vector based on generative adversarial networks and differential interpolation. The specific steps are as follows: Generative Adversarial Networks (GANs) are constructed based on time series data. The time series feature vectors are input into the GAN, and adversarial samples are generated based on the adversarial game process between the generator and the discriminator of the GAN. The adversarial examples are completed using a differential interpolation algorithm to generate a preprocessed historical payload dataset. S2. Input the calibration set into the MoE model that has been pre-trained using the training set to perform load prediction and output the load prediction value. The MoE model comprises an expert layer, a gated routing layer, and a prediction layer, specifically: S21. The expert layer comprises multiple convolutional attention sub-networks based on contrastive learning as expert networks. Each expert network includes a one-dimensional convolutional layer and a Transformer layer containing a multi-head self-attention mechanism, wherein: the one-dimensional convolutional layer is used to extract the mutation features of the preprocessed historical payload dataset, calculated as follows: ; In the formula, Indicates mutation characteristics; express function; This represents a one-dimensional convolutional layer; This represents the preprocessed historical load data sequence as input. Mapping mutation features to a query matrix of a multi-head self-attention mechanism Key matrix Sum matrix ; The contextual dependencies of mutation features are established based on a multi-head self-attention mechanism, resulting in a feature matrix, which is calculated as follows: ; In the formula, Represents the characteristic matrix; express function; Represents the preset projection vector; Indicates the transpose operation; The feature matrix is ​​passed through a feedforward neural network layer to output the initial prediction value corresponding to each expert network. S22. The gated routing layer is constructed based on a multilayer perceptron. It extracts multi-site meteorological covariates from the expanded historical load dataset and fuses them with the acquired static features to obtain the gated input. The routing probability distribution of the gated input on each expert network is calculated using the Softmax function. The calculation method is as follows: ; In the formula, Indicates the first Routing probability distribution of an expert network; Indicates the first The weights of an expert network; Indicates the first Bias of an expert network; Indicates gating input; Indicates the first The weights of an expert network; Indicates the first Bias of an expert network; Indicates the number of expert networks; The static features refer to features that are independent of time, such as which industry (large industry, commerce, residential) this monitoring node belongs to, or its geographical latitude and longitude. S23, the prediction layer includes two quantile prediction heads and a Top function. k-sparse activation mechanism; Two high and low quantile prediction heads are set side by side in the prediction layer. Each prediction head corresponds to a specific quantile, such as 0.1 and 0.9. The lower quantile of 0.1 is the lower bound, and the higher quantile of 0.9 is the upper bound. Load forecasts are calculated based on the initial forecasts and routing probability distributions of each expert network, including upper quantile load forecasts and lower quantile load forecasts. The calculation method is as follows: ; ; In the formula, This represents the lower quantile load forecast value; This represents the upper quantile load forecast value; Indicates the first The initial predictions of the low quantile of an expert network; Indicates the first The initial predictions of the high quantile of an expert network; Indicates the number of active expert networks; S24. The loss function of the MoE model is a load balancing loss function, used to balance the probability of route allocation to each expert network. The calculation method is as follows: ; In the formula, Represents the load balancing loss function; This represents the preset weight penalty hyperparameter; This indicates the number of trainees assigned to the current training batch. The frequency percentage of each expert network; This indicates that the first [number] trainer in the current training batch is selected. The expected routing probability of an expert network; S25, also includes introducing a dynamic programming algorithm as a distance metric for the contrastive learning loss mechanism during the pre-training phase, specifically: Positive and negative sample pairs are constructed based on the preprocessed historical load dataset. Sample pairs belonging to the same specific fluctuation pattern (such as both belonging to peak loads caused by meteorological changes) are defined as positive sample pairs, and sample pairs belonging to different fluctuation patterns are defined as negative sample pairs. Extract the feature matrices of any two expert networks; The dynamic programming algorithm is used to calculate the dynamic time programming distance between two feature matrices. The calculation method is as follows: ; In the formula, Representation of the characteristic matrix and characteristic matrix Dynamic time planning distance between them; Describes the minimum value function; express Feature points and Feature points The Euclidean distance between them; To address the issue that power load curves are prone to local shifts or scaling on the time axis, dynamic programming algorithms can break the strict time alignment constraints and accurately measure the similarity of loads in macroscopic shape and evolution trend. Using this as a measure for contrastive learning can induce the parallel expert networks in the MoE model to spontaneously converge to a specific power physical form, solving the problems of homogenization of expert responsibilities and feature collapse in traditional MoE models. S3. Determine the upper and lower limits of the load interval based on the load forecast value, specifically: take the upper quantile load forecast value as the upper limit of the load interval and the lower quantile load forecast value as the lower limit of the load interval. S4. Calculate the prediction boundary residual sequence using the upper and lower bounds of the load interval. The prediction boundary refers to the upper and lower bounds calculated by the model itself. The difference between the actual value and the boundary is the prediction boundary residual. Obtain the current load data and actual load observations; Calculate the difference between the lower bound of the load interval corresponding to the load data at the current moment and the actual load observation value, and use it as the first candidate prediction boundary residual; Calculate the difference between the actual load observation value and the upper bound of the load interval corresponding to the load data at the current time, and use it as the second candidate prediction boundary residual; The maximum value between the first candidate prediction boundary residual and the second candidate prediction boundary residual is selected as the prediction boundary residual at the current time. The specific calculation method is as follows: ; In the formula, express The predicted boundary residual at time step; express Load data at any given time; express The actual load observation value at any given time; Sort the predicted boundary residuals at all times in ascending order to obtain the predicted boundary residual sequence; S5. Constructing long and short time windows specifically involves setting their size. Short-term sliding window, size is The long-term sliding window and the corresponding long-term and short-term forgetting factors; S6. Calculate the corresponding long-short-term boundary weights based on the long-short-term forgetting factor. The long-short forgetting factor is calculated by multiplying the exponential decay term of the current time-to-time long-short forgetting factor by the sum of the exponential decay terms of all times within a preset time period. The long-short boundary weights are obtained by calculating the ratio of the exponential decay term of the long-short forgetting factor at the current time to the sum of the exponential decay terms at all times within the preset time period. The calculation method is as follows: , ; , ; In the formula, express Short-term boundary weights at time points; express Short-term forgetting factor; express Short-term forgetting factor; Indicates the total time; Indicates a short-term sliding window; Indicates a long-term sliding window; express Long-term boundary weights at each time step; express Long-term forgetting factors over time; express Long-term forgetting factors over time; S7. Obtain the short- and long-term critical prediction boundary residuals based on the preset target fault tolerance rate and short- and long-term boundary weights. The long and short-term boundary weights corresponding to the predicted boundary residual sequence are accumulated until the accumulated weight first reaches or exceeds the target coverage, at which point the accumulation stops. The target coverage is the difference between 1 and the target fault tolerance rate, calculated as follows: ; ; In the formula, Indicates the first A short-term boundary weight; Indicates the first A long-term boundary weight; This indicates the preset target fault tolerance rate; An index representing the short-term boundary weights; An index representing the long-term boundary weight; S8. Extract long and short-term prediction boundary residuals from the prediction boundary residual sequence based on the index of long and short-term critical prediction boundary residuals. Extract the index of the last accumulated short-term boundary weight. ; Extract the index of the last accumulated long-term boundary weight. ; The first The prediction boundary residuals corresponding to each boundary weight are used as the short-term prediction boundary residuals. The first The prediction boundary residuals corresponding to each boundary weight are used as the long-term prediction boundary residuals. S9. The global safety margin is calculated based on the long-term and short-term prediction boundary residuals. The global safety margin is obtained by calculating the short-term and long-term prediction boundary residuals based on adaptive fusion weighting. The calculation method is as follows: ; In the formula, express Global safety margin at any given moment; express Adaptive fusion weights at each time step; This represents the short-term forecast boundary residual; This represents the long-term forecast boundary residual; The adaptive fusion weight is calculated based on a preset sensitivity coefficient, the current load fluctuation intensity index, and a preset fluctuation threshold. The calculation method is as follows: ; In the formula, Represents a mapping function; Indicates the sensitivity coefficient; express The load fluctuation intensity index at any given time is used to quantify the degree of drastic change in power load near the current moment. A value approaching 1 indicates that the system is dominated by a short-term mutation window; when A value approaching 0 indicates that the system is dominated by a long-term trend window. Indicates the preset threshold for fluctuation; In this embodiment, the load fluctuation intensity index is calculated as follows: Take the continuous load values ​​within a preset time period before the current moment, calculate the ratio of its standard deviation to the mean, and map the ratio through a mapping function to obtain the load fluctuation intensity index; Take the load values ​​from the past 8 time points as [120, 122, 118, 135, 142, 138, 145, 150]; The calculated mean is 133.75, the calculated standard deviation is 11.23, and the calculated ratio is 0.084. Using the tanh function for mapping, the load fluctuation intensity index is obtained as 0.0838; S10. Combining the global safety margin and the upper and lower limits of the load interval, the power load interval is obtained, denoted as: ; In one specific embodiment, load data is collected every half hour throughout a 24-hour period on a certain day, and the target fault tolerance rate for power grid dispatch is set. =0.1; The historical load dataset is input into the trained MoE model, and the upper quantile load prediction is output. =135.2MW and lower quantile load forecast =120.5MW; Obtain the actual load observation value at the current moment. =136.0MW; the predicted boundary residual sequence is calculated using the upper and lower boundaries of the load interval, for example... MW; Calculate the predicted boundary residuals at each time step and sort them in ascending order to obtain... ; Assuming the current time is 14:00, set... The short-term time sliding window is set to 5, representing the five most recent historical moments: 13:00, 12:30, 12:00, 11:30, and 11:00, with a short-term forgetting factor of 0.5. set up The value is 10, representing the 10 most recent historical moments: a long-term sliding window from 13:00 to 09:30, with a long-term forgetting factor of 0.95; The short-term boundary weights for each time point are calculated, for example, the five time points between 11:00 and 13:00. ; The short-term boundary weights are accumulated, i.e. ,therefore Take the 4th one. =2.9MW, as ; The long-term boundary weights for each time point are calculated, for example, for 10 time points between 11:00 and 13:00. ; The long-term boundary weights are accumulated, i.e. ,therefore Take the 9th one. =2.5MW, as ; The sensitivity coefficient is set to 5, the load fluctuation intensity index at the current moment is 0.85, and the fluctuation preset threshold is 0.6. The mapping function adopts the Sigmoid function, and the adaptive fusion weight is calculated to be 0.777. Calculate the global safety margin =0.777×2.9+0.223×2.5=2.2533+0.5575=2.8108 MW; The final power load range is =[117.69,138.01]MW.

[0024] Example 2: This embodiment provides a dynamic power load interval prediction system based on long and short window periods. The system includes: The data acquisition module acquires historical load datasets of target monitoring nodes in the power system and preprocesses them, dividing the preprocessed historical load datasets into training and calibration sets according to a preset ratio. The load boundary prediction module takes the calibration set as input to the hybrid expert MoE model pre-trained with the training set to perform load prediction and outputs the load prediction value; it determines the upper and lower bounds of the load interval based on the load prediction value and uses the upper and lower bounds of the load interval to calculate the prediction boundary residual sequence. The long-short-term prediction boundary residual calculation module constructs long-short-term time windows and sets corresponding long-short-term forgetting factors, calculates the corresponding long-short-term boundary weights based on the long-short-term forgetting factors, and obtains the long-short-term critical prediction boundary residuals based on the preset target fault tolerance rate and long-short-term boundary weights. Extract long and short-term prediction boundary residuals from the prediction boundary residual sequence based on the index of long and short-term critical prediction boundary residuals. The power load interval calculation module uses the long-term and short-term forecast boundary residuals to calculate the global safety margin, and obtains the power load interval based on the global safety margin and the upper and lower boundaries of the load interval.

[0025] Example 3: This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements a dynamic power load interval prediction method based on long and short windows as described in any one of Embodiment 1.

[0026] Example 4: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a dynamic power load interval prediction method based on a long-short window as described in any one of Embodiment 1.

[0027] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, A and B simultaneously, or B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, at least one of a, b, and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.

[0028] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0029] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0030] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0031] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A dynamic power load interval prediction method based on long and short window, characterized in that, The method includes: Acquire historical load datasets of target monitoring nodes in the power system and preprocess them. Divide the preprocessed historical load datasets into training sets and calibration sets according to a preset ratio. The calibration set is input into the hybrid expert MoE model pre-trained with the training set to predict the load and output the predicted load value. The upper and lower bounds of the load interval are determined based on the predicted load value, and the prediction boundary residual sequence is calculated using the upper and lower bounds of the load interval. Construct long and short time windows and set corresponding long and short forgetting factors. Calculate the corresponding long and short boundary weights based on the long and short forgetting factors. Obtain the long and short critical prediction boundary residuals based on the preset target tolerance rate and long and short boundary weights. Based on the index of the long- and short-term critical prediction boundary residuals, the long- and short-term prediction boundary residuals are extracted from the prediction boundary residual sequence. The global safety margin is calculated using the long- and short-term prediction boundary residuals. The power load range is obtained based on the global safety margin and the upper and lower boundaries of the load range.

2. The dynamic power load interval prediction method based on long and short window as described in claim 1, characterized in that, The historical load dataset includes historical load observation sequences, multi-site environmental meteorological covariates, and time attribute labels. The preprocessing includes label mapping, feature construction, and long-tailed distribution sample augmentation, wherein: Specifically, the label mapping uses a trigonometric function encoding mechanism to map time attribute labels to historical load observation sequences, in order to preserve the periodicity and continuity of historical load observation sequences. The feature construction specifically involves extracting load statistics parameters based on a preset sliding window mechanism, and then concatenating them with the acquired historical lagged load observation sequences and historical multi-site environmental meteorological covariates to construct a time series feature vector. The long-tailed distribution sample augmentation of time series feature vectors based on generative adversarial networks and differential interpolation is performed as follows: Generative Adversarial Networks (GANs) are constructed based on time series data. The time series feature vectors are input into the GAN, and adversarial samples are generated based on the adversarial game process between the generator and the discriminator of the GAN. The adversarial examples are completed using a differential interpolation algorithm to generate a preprocessed historical payload dataset.

3. The dynamic power load interval prediction method based on long and short-term windows according to claim 2, characterized in that, The MoE model comprises an expert layer, a gated routing layer, and a prediction layer, specifically: The expert layer includes multiple convolutional attention sub-networks based on contrastive learning as expert networks. Each expert network includes a one-dimensional convolutional layer and a Transformer layer containing a multi-head self-attention mechanism. Specifically, the one-dimensional convolutional layer is used to extract mutation features from the preprocessed historical load dataset, the contextual dependency of mutation features is established based on the Transformer layer containing a multi-head self-attention mechanism, and the initial prediction value corresponding to each expert network is output. The gated routing layer is constructed based on a multilayer perceptron. It extracts multi-site meteorological covariates from the preprocessed historical load dataset and fuses them with the acquired static features to obtain the gated input. The routing probability distribution of the gated input on each expert network is calculated based on the Softmax function. The prediction layer includes two quantile prediction heads and a Topology execution layer. The k-sparse activation mechanism calculates load forecasts based on the initial forecasts and routing probability distributions of each expert network, including upper quantile load forecasts and lower quantile load forecasts. The upper quantile load forecast is used as the upper limit of the load interval, and the lower quantile load forecast is used as the lower limit of the load interval.

4. The dynamic power load interval prediction method based on long and short window as described in claim 1, characterized in that, The prediction boundary residual sequence is calculated using the upper and lower bounds of the load interval, specifically as follows: Obtain the current load data and actual load observations; Calculate the difference between the lower bound of the load interval corresponding to the load data at the current moment and the actual load observation value, and use it as the first candidate prediction boundary residual; Calculate the difference between the actual load observation value and the upper bound of the load interval corresponding to the load data at the current time, and use it as the second candidate prediction boundary residual; The maximum value between the first candidate prediction boundary residual and the second candidate prediction boundary residual is selected as the prediction boundary residual at the current time. Sort the predicted boundary residuals at all times in ascending order to obtain the predicted boundary residual sequence.

5. The dynamic power load interval prediction method based on long and short window as described in claim 1, characterized in that, Constructing long and short time windows specifically involves setting the size to... Short-term sliding window, size is The long-term sliding window and the corresponding long-term and short-term forgetting factors; The long-short boundary weights are obtained by calculating the ratio of the exponential decay term of the long-short forgetting factor at the current moment to the sum of the exponential decay terms at all moments within the preset time period.

6. The dynamic power load interval prediction method based on long and short window as described in claim 1, characterized in that, Based on the preset target fault tolerance rate and long- and short-term boundary weights, the long- and short-term critical prediction boundary residuals are obtained as follows: The long and short-term boundary weights corresponding to the predicted boundary residual sequence are accumulated until the accumulated weight first reaches or exceeds the target coverage, at which point the accumulation stops. The target coverage is the difference between 1 and the target tolerance. Extract the index of the last accumulated short-term boundary weight. ; Extract the index of the last accumulated long-term boundary weight. ; The first The prediction boundary residuals corresponding to each boundary weight are used as the short-term prediction boundary residuals. The first The prediction boundary residuals corresponding to each boundary weight are used as the long-term prediction boundary residuals. The global safety margin is obtained by calculating the short-term and long-term prediction boundary residuals based on adaptive fusion weighting.

7. The dynamic power load interval prediction method based on long and short window as described in claim 6, characterized in that, The adaptive fusion weight is calculated based on a preset sensitivity coefficient, the load fluctuation intensity index at the current moment, and a preset fluctuation threshold.

8. A dynamic power load interval prediction system based on long and short window, characterized in that, The system includes: The data acquisition module acquires historical load datasets of target monitoring nodes in the power system and preprocesses them, dividing the preprocessed historical load datasets into training and calibration sets according to a preset ratio. The load boundary prediction module takes the calibration set as input to the hybrid expert MoE model pre-trained with the training set to perform load prediction and outputs the load prediction value; it determines the upper and lower bounds of the load interval based on the load prediction value and uses the upper and lower bounds of the load interval to calculate the prediction boundary residual sequence. The long-short-term prediction boundary residual calculation module constructs long-short-term time windows and sets corresponding long-short-term forgetting factors, calculates the corresponding long-short-term boundary weights based on the long-short-term forgetting factors, and obtains the long-short-term critical prediction boundary residuals based on the preset target fault tolerance rate and long-short-term boundary weights. Extract long and short-term prediction boundary residuals from the prediction boundary residual sequence based on the index of long and short-term critical prediction boundary residuals. The power load interval calculation module uses the long-term and short-term forecast boundary residuals to calculate the global safety margin, and obtains the power load interval based on the global safety margin and the upper and lower boundaries of the load interval.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images