Power load forecasting method, device, computing equipment and storage medium for extreme weather events

Abstract

The present application provides a method, device, computing equipment and storage medium for power load prediction of extreme weather events, wherein the method comprises: collecting power load data and meteorological forecast data; determining the correlation between power load data and meteorological forecast data based on the correlation coefficient method, and eliminating redundant features to obtain a load data sequence; processing the load data sequence through variable mode decomposition optimized by a particle swarm algorithm to obtain high-frequency components, low-frequency components and residual components; processing the high-frequency components through a first model to obtain a first sequence, and processing the low-frequency components through a second model to obtain a second sequence; processing the residual components through a third model, and combining the processing results with the first sequence and the second sequence to determine the power load prediction data. Under the premise of reducing feature redundancy and information calculation, high-reliability signal decomposition is used to reduce the complexity of data processing and complete complex extreme weather prediction tasks.

Description

Technical Field

[0001] The present application relates to the technical field of power load forecasting, and in particular to a power load forecasting method for extreme weather events. The present application also relates to a power load forecasting device for extreme weather events, a computing device, and a computer-readable storage medium. Background Art

[0002] At present, extreme weather events are increasing and intensifying around the world, manifested in persistent high temperatures, droughts, snowstorms, and more frequent extreme climate events such as heat waves and cold waves. Extreme weather is characterized by low probability, high risk, and great harm, often leading to sharp changes in electricity demand or fluctuations in renewable energy output, posing a major challenge to the normal operation and power supply of wind and photovoltaic power generation systems. In addition, extreme weather events may cause significant damage to infrastructure and immediately increase or decrease electricity demand, seriously affecting the stable power supply and dispatching decisions of the power grid.

[0003] Electricity load forecasting has formed a relatively complete technical system, in which the combined model can overcome the defects of a single model and achieve complementary advantages, so it is widely used. However, the current research on load forecasting under extreme weather conditions is relatively limited. Most studies mainly consider weather factors as characteristic variables, but lack research on specific weather forecasting scenarios. The impact mechanism of heat waves, cold waves and normal weather on load is not consistent. Cold waves cause transmission lines to break due to freezing or collapse due to heavy snow and ice, which in turn affects the normal supply of load. Under continuous high temperature conditions, the air conditioning load in industrial areas has surged.

[0004] Due to extreme weather, the power load sequence is highly volatile and mutated, which includes not only trend and periodic components, but also a large amount of white noise, which seriously affects the accuracy of power load forecasting. For this reason, signal decomposition technology is widely used in the field of power load forecasting. Rubasinghe et al. adopted the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) technology to eliminate the high-frequency components of the original net load, and then input the long short-term memory (LSTM) network prediction to obtain accurate day-ahead prediction. However, when processing complex signals, especially when the frequencies of signal components are close or overlapping, modal aliasing may still occur, resulting in large errors in prediction. Mokarram et al. introduced a technology that combines multi-input LSTM, fuzzy system and discrete wavelet transform. Through discrete wavelet transform, the technology decomposes the net load into components of different frequencies, which show significant localization characteristics in time and frequency. In addition, this method can also reveal potential details, which helps to predict highly uncertain points in complex systems. However, due to the strong load volatility under extreme high and low temperatures, it is difficult for existing single prediction models to accurately predict load trends. In addition, using a single decomposition model with manual tuning cannot effectively select hyperparameters, resulting in poor decomposition performance.

[0005] In order to more accurately predict electricity demand, researchers have proposed various load forecasting methods. Generally, short-term load forecasting models can be divided into three categories: statistical models, deep learning models, and combined models. Each method has its own unique advantages and is suitable for handling different forecasting scenarios and needs. Statistical methods rely on mathematical analysis of historical data and emphasize the stability and explanatory power of the model, such as linear regression, exponential smoothing, and autoregressive integrated moving average; deep learning models tend to learn complex patterns and associations from data, which improves the accuracy and flexibility of predictions. In particular, convolutional neural networks, deep belief networks, gated recurrent units, long short-term memory, and recurrent neural networks have been widely used. However, statistical methods and single deep learning methods often show poor prediction performance when facing complex extreme weather scenarios and sparse data scenarios. Most traditional prediction methods have poor nonlinear time dependency extraction capabilities and relatively single feature extraction methods, which are limited to extracting time features and extracting spatial features. They cannot effectively and comprehensively learn more time information in the frequency domain, and cannot effectively obtain information between meteorological features and load features, thus performing poorly in complex extreme weather prediction tasks. Summary of the invention

[0006] In view of this, the embodiment of the present application provides a method for predicting power load in extreme weather events to solve the technical defects in the prior art. The embodiment of the present application also provides a device for predicting power load in extreme weather events, a computing device, and a computer-readable storage medium.

[0007] According to a first aspect of an embodiment of the present application, a method for predicting power load for extreme weather events is provided, comprising:

[0008] Collect power load data and weather forecast data;

[0009] Based on the correlation coefficient method, the correlation between the power load data and the weather forecast data is determined, and redundant features are eliminated to obtain a load data sequence;

[0010] Processing the load data sequence by variable mode decomposition optimized by particle swarm algorithm to obtain high-frequency components, low-frequency components and residual components;

[0011] The high-frequency component is processed by a pre-trained first model to obtain a first sequence, and the low-frequency component is processed by a pre-trained second model to obtain a second sequence;

[0012] The residual component is processed by a pre-trained third model, and the processing result is combined with the first sequence and the second sequence to determine the power load forecast data.

[0013] Optionally, it also includes:

[0014] Filling missing values ​​of the power load data based on linear interpolation and polynomial interpolation;

[0015] Based on the standard score and Hakone plot method, outliers of the power load data are processed.

[0016] Optionally, the processing the high frequency component by using a pre-trained first model to obtain a first sequence includes:

[0017] Extracting features of the high-frequency components through the temporal convolutional network included in the first model, and fusing the feature extraction results with features corresponding to the weather forecast data to obtain input features;

[0018] Along the channel dimension of the first model, the input feature is divided into a plurality of subsequences by a frequency domain attention mechanism, and a frequency channel vector is determined by the plurality of subsequences;

[0019] The frequency channel vector is processed according to the gated pulse neural P model included in the first model to obtain the first sequence.

[0020] Optionally, the processing the low-frequency component by using a pre-trained second model to obtain a second sequence includes:

[0021] The second model based on the iTransfomer model architecture includes a normalization layer, a feedforward network and a self-attention mechanism to process the low-frequency component to obtain the second sequence.

[0022] Optionally, the processing of the residual component by a pre-trained third model and combining the processing result with the first sequence and the second sequence to determine the power load forecast data includes:

[0023] According to the multi-head sparse self-attention mechanism, the decoder and the Informer model distillation mechanism included in the third model of the Informer model architecture, the residual component is processed to obtain residual load component prediction data;

[0024] The power load forecast data is determined by combining the first sequence, the second sequence and the residual load component forecast data.

[0025] Optionally, the process of determining the variable mode decomposition is:

[0026] Based on Hilbert transform, determining analytical signals of all modal functions, and according to all the analytical signals, determining the single-sided spectra and center frequencies of all the analytical signals;

[0027] According to the center frequency, the single-sided spectrum is modulated to a corresponding baseband, and a constrained variable mode decomposition problem is constructed;

[0028] Based on the introduced quadratic penalty factor, the constrained variable mode decomposition problem is transformed into an unconstrained variable mode decomposition problem, and an extended Lagrangian expression is obtained;

[0029] The Lagrangian expression is gradually optimized by an alternating direction multiplier iteration algorithm to determine the variable mode decomposition.

[0030] Optionally, the process of optimizing the variable mode decomposition by the particle swarm algorithm includes:

[0031] Initializing a potential solution of the variable mode decomposition into a plurality of particles;

[0032] Calculating the historical optimal position of any of the particles, and recording the historical optimal positions of all the particles;

[0033] Based on the historical optimal positions of all the particles, determining the individual historical optimal fitness value of any particle and the group historical optimal fitness value of all the particles;

[0034] Based on the individual historical optimal fitness value and the group historical optimal fitness value, updating the positions and velocities of all the particles;

[0035] Repeat the steps of determining the individual historical optimal fitness value of any of the particles and the group historical optimal fitness value of all the particles until the maximum number of iterations is reached, and obtain the variable mode decomposition optimized by the particle swarm algorithm.

[0036] According to a second aspect of an embodiment of the present application, there is provided a device for predicting power load for extreme weather events, comprising:

[0037] A collection module configured to collect power load data and weather forecast data;

[0038] A removal module is configured to determine the correlation between the power load data and the weather forecast data based on a correlation coefficient method, and remove redundant features to obtain a load data sequence;

[0039] A decomposition module is configured to process the load data sequence through variable mode decomposition optimized by a particle swarm algorithm to obtain a high-frequency component, a low-frequency component and a residual component;

[0040] A first processing module is configured to process the high-frequency component through a pre-trained first model to obtain a first sequence, and to process the low-frequency component through a pre-trained second model to obtain a second sequence;

[0041] The second processing module is configured to process the residual component through a pre-trained third model, and combine the processing result with the first sequence and the second sequence to determine the power load prediction data.

[0042] According to a third aspect of an embodiment of the present application, a computing device is provided, including:

[0043] Memory and processor;

[0044] The memory is used to store computer-executable instructions, and the processor can implement the steps of the method for predicting power load for extreme weather events when executing the computer-executable instructions.

[0045] According to a fourth aspect of an embodiment of the present application, a computer-readable storage medium is provided, which stores computer-executable instructions, and when the instructions are executed by a processor, the steps of the method for predicting power load for extreme weather events are implemented.

[0046] According to a fifth aspect of an embodiment of the present application, a chip is provided, which stores a computer program, and when the computer program is executed by the chip, the steps of the method for predicting power load for extreme weather events are implemented.

[0047] The power load prediction method for extreme weather events provided in the present application is to collect power load data and meteorological forecast data; determine the correlation between the power load data and the meteorological forecast data based on the correlation coefficient method, and eliminate redundant features to obtain a load data sequence; process the load data sequence through variable mode decomposition optimized by particle swarm algorithm to obtain high-frequency components, low-frequency components and residual components; process the high-frequency components through a pre-trained first model to obtain a first sequence, and process the low-frequency components through a pre-trained second model to obtain a second sequence; process the residual components through a pre-trained third model, and combine the processing results with the first sequence and the second sequence to determine the power load prediction data. Under the premise of reducing feature redundancy and information calculation, with high-reliability signal decomposition, the complexity of data processing is weakened to complete complex extreme weather prediction tasks, and further by accurately predicting the load under extreme weather, it effectively assists power grid operators in optimizing resource allocation, thereby ensuring the stable operation of the power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0049] Figure 1 This is a flow chart of a method for predicting power load for extreme weather events provided by an embodiment of the present application;

[0050] Figure 2 It is a schematic diagram of a framework of a method for predicting power load for extreme weather events provided in one embodiment of the present application;

[0051] Figure 3 It is a schematic diagram of a FAT-FAGP model of a power load forecasting method for extreme weather events provided in one embodiment of the present application;

[0052] Figure 4 It is a schematic diagram of an iTransformer model of a power load forecasting method for extreme weather events provided in one embodiment of the present application;

[0053] Figure 5 It is a schematic diagram of an Informer model of a power load forecasting method for extreme weather events provided in one embodiment of the present application;

[0054] Figure 6 It is a structural schematic diagram of a power load prediction device for extreme weather events provided by an embodiment of the present application;

[0055] Figure 7 It is a structural block diagram of a computing device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0056] Many specific details are described in the following description to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present application, so the present application is not limited by the specific implementation disclosed below.

[0057] The terms used in one or more embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit one or more embodiments of the present application. The singular forms of "a", "said" and "the" used in one or more embodiments of the present application and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used in one or more embodiments of the present application refers to and includes any or all possible combinations of one or more associated listed items.

[0058] It should be understood that, although the terms first, second, etc. may be used to describe various information in one or more embodiments of the present application, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of the present application, the first may also be referred to as the second, and similarly, the second may also be referred to as the first.

[0059] In the present application, a method for predicting power load in extreme weather events is provided. The present application also relates to a device for predicting power load in extreme weather events, a computing device, and a computer-readable storage medium, which are described in detail one by one in the following embodiments.

[0060] Figure 1 A flowchart of a method for predicting power load for extreme weather events provided in accordance with an embodiment of the present application is shown, which specifically includes the following steps:

[0061] Step S102: Collecting power load data and weather forecast data;

[0062] Step S104: determining the correlation between the power load data and the weather forecast data based on the correlation coefficient method, and removing redundant features to obtain a load data sequence;

[0063] Step S106: Processing the load data sequence by variable mode decomposition optimized by particle swarm algorithm to obtain high-frequency components, low-frequency components and residual components;

[0064] Step S108: Processing the high-frequency component by a pre-trained first model to obtain a first sequence, and processing the low-frequency component by a pre-trained second model to obtain a second sequence;

[0065] Step S110: Process the residual component through a pre-trained third model, and combine the processing result with the first sequence and the second sequence to determine power load forecast data.

[0066] Among them, for the collected power load data and weather forecast data, the correlation between the two is first determined, and the variables contained in the weather forecast data are clarified. There is a strong correlation between the load values ​​in the power load data, and then the redundant features that will not have a significant impact on the load value are eliminated to obtain a simplified model of the load data series and improve the prediction accuracy; and the weather forecast data includes temperature, precipitation, humidity and other data.

[0067] Based on this, variable mode decomposition is a time-frequency signal processing method that can change the scale. It overcomes the shortcomings of mode aliasing and endpoint effect in the adaptive decomposition algorithm. The essence of variable mode classification is Wiener filtering denoising, which has a good denoising effect. However, with the widespread application of decomposition methods, a single decomposition method has the problem of falling into the local optimality of hyperparameters, resulting in incomplete decomposition or too many modes. In addition, due to the non-stationary and nonlinear characteristics of the load, variable mode decomposition is optimized by particle swarm algorithm to reduce the complexity of the original sequence. The optimized variable mode decomposition decomposes the complex load signal into high-frequency components, low-frequency components and residual components, providing a solid modeling foundation for prediction.

[0068] Afterwards, the decomposed high-frequency components, low-frequency components and residual components are processed using the corresponding first model, second model and third model respectively, and all processing results are reconstructed to obtain the final prediction result, that is, the power load forecast data, to predict the load data for the next few months or days. The time granularity is in minutes, such as 15 minutes, and the final result presents the prediction results in the form of charts, including historical data comparison and future predictions. The output covers real-time weather forecast data, load forecast results and evaluation indicators, and supports multiple result export formats, including txt, CSV, Excel and database connection.

[0069] In addition, the output results can select an appropriate time range to display the forecast data, such as the next 24 hours, 7 days or 30 days. According to actual needs, it may also be necessary to provide users with the function of selecting a time range. Users can specify the time period by dragging the slider or selecting the start and end dates, mark important time nodes on the timeline, such as the start time of the forecast, the time when key events occur, etc., and clearly display the values ​​of key forecast results, such as the expected maximum load, minimum load, average load, etc.

[0070] Different types such as power demand and load peak are also set up for filtering forecast results. Multiple types can be selected at the same time for comprehensive analysis. Data charts and statistical information are dynamically updated according to the selected result type. The data level selection function is added to realize the selection of charts showing different data levels or time periods.

[0071] It also allows custom load thresholds, triggering an alarm when the actual load exceeds or falls below the set value. It supports setting alarms of different levels, such as high, medium, and low, and triggering different notification types based on the threshold level. It notifies users of the specific value, time, impact range, and other information when the load exceeds the threshold through pop-up reminders, email notifications, SMS alerts, and other notification methods.

[0072] Furthermore, after collecting the power load data and the weather forecast data, since there are a large number of missing and abnormal values ​​in the collected power load data, which will affect the subsequent prediction accuracy, it is necessary to pre-process the power load data. In this embodiment, the specific implementation method is as follows:

[0073] Based on linear interpolation and polynomial interpolation, missing values ​​of the power load data are filled; based on standard score and Hakken plot method, outliers of the power load data are processed.

[0074] Among them, Figure 2 As shown in the framework diagram of a method for power load prediction of extreme weather events, the collected power load data and weather forecast data need to be subjected to correlation analysis after data preprocessing.

[0075] Specifically, the Kendall correlation coefficient method can be used to screen out the most relevant variable combination and remove redundant features. The Kendall correlation coefficient is a statistical indicator used to measure the rank correlation between two random variables, and is usually used to evaluate the nonlinear relationship between variables. In the Kendall test, the calculated correlation coefficient is used to test the statistical dependence between two random variables, without the need to make parameter assumptions on the data.

[0076] The most commonly used methods are Kendall Tau-b and Kendall-Tau-c correlation coefficients. The Kendall-Tau correlation coefficient is suitable for determining the correlation between two columns of discrete ordered data. Taking the application of the Kendall-Tau-b correlation coefficient in load forecasting under extreme weather conditions as an example, the formula for Tau-b is as follows:

[0077] ,

[0078] Among them, C represents the number of consistent pairs; D represents the number of inconsistent pairs; N represents the number of samples; Indicates the number of elements contained in the i-th small set; Represents the number of elements contained in the jth small set.

[0079] For any type of meteorological forecast data, such as high temperature, drought, and blizzard, the corresponding load data sequence can be selected by the correlation coefficient method, that is, Figure 2 In addition, for any type of extreme weather corresponding to the meteorological forecast data, such as Figure 2 The obtained load data series also needs to be normalized to ensure data consistency.

[0080] In summary, the discretely sorted power load data and weather forecast data are combined to obtain n data combinations. They can be divided into three categories, namely consistent pairs, inconsistent pairs, and unchanged pairs. If the power load data and a certain weather forecast feature are positively correlated, the two variables change in the same direction, and such a pair is called a consistent pair. If the power load data and a certain weather forecast feature are negatively correlated, the two variables change in the direction of the inconsistent pair, and such a pair is called an inconsistent pair. If the number of consistent pairs is significantly greater than the number of inconsistent pairs, it indicates that the power load data is positively correlated with a certain weather forecast feature; otherwise, there is a negative correlation. Repeat this process until all features are involved in the calculation.

[0081] Screening out the optimal feature coupling variable combination can effectively reduce information loss and improve prediction performance. In order to determine which variables may have an impact on load forecasting, the Kendall correlation coefficient method is used to filter out the most relevant variable combination and remove redundant features, thereby simplifying the model and improving prediction accuracy, and providing a better basis for subsequent modeling. Finally, data normalization is used to reduce the differences between different features.

[0082] Furthermore, the implementation of variable mode decomposition, in this embodiment, is specifically implemented as follows:

[0083] Based on Hilbert transform, the analytical signals of all modal functions are determined, and according to all the analytical signals, the single-sided spectra and the center frequencies of all the analytical signals are determined; according to the center frequency, the single-sided spectrum is modulated to the corresponding baseband to construct a constrained variable mode decomposition problem; based on the introduced quadratic penalty factor, the constrained variable mode decomposition problem is converted into an unconstrained variable mode decomposition problem to obtain an extended Lagrangian expression; the Lagrangian expression is gradually optimized through an alternating direction multiplier iteration algorithm to determine the variable mode decomposition.

[0084] Among them, the main purpose of variable mode decomposition is to decompose the original signal into multiple intrinsic mode functions, each of which exhibits a relatively smooth spectral characteristic. For its core variational problem (variable mode decomposition problem), its goal is to minimize the estimated bandwidth and bandwidth of all modes. In this framework, each intrinsic mode function is considered to be a finite bandwidth signal with a specific center frequency. To solve this variational problem, the alternating direction multiplier method is used. This method continuously iterates and updates each mode and its center frequency, gradually adjusting to their respective basebands, thereby extracting all modes and their corresponding center frequencies. The size of the mode number has an important influence on the decomposition effect of the variable mode decomposition algorithm. If the mode value is too large, it is easy to produce mode aliasing or generate additional noise; if the mode number is too small, the modal decomposition will be insufficient.

[0085] Specifically, if f represents the load sequence after preprocessing, f The decomposition steps are as follows:

[0086] First, the variable mode decomposition problem is constructed, and based on the Hilbert transform, each mode function is obtained The analytical signal of , the single-sided spectrum is as follows:

[0087] ,

[0088] Mix the analytical signals of each mode and estimate the center frequency , modulate the spectrum of each mode to the corresponding baseband, where the baseband is as follows:

[0089] ,

[0090] Calculate the square L2 norm of the above demodulated signal gradient and estimate the bandwidth of each modal signal. The constrained variable mode decomposition problem is as follows:

[0091] ,

[0092] in, and Respectively represent k modal components and their center frequencies; K Indicates the modal number obtained by decomposition; represents the Dirac function.

[0093] After that, the variable mode decomposition problem is solved. First, the constrained variable mode decomposition problem is transformed into an unconstrained form. The introduced quadratic penalty factor still ensures the high aggregation accuracy of the signal under the influence of Gaussian noise. Lagrangian operator It is used to maintain the strict execution of the constraints, which is reflected by the following extended Lagrangian expression:

[0094] ,

[0095] Finally, the objective function is gradually optimized using the alternating direction multiplier iteration algorithm, and the frequency center and modal function are iteratively updated until convergence and the specified number of modal components VIMF are obtained to determine the variable modal decomposition. The variable modal decomposition technology converts the signal decomposition process into a non-recursive, variational modal decomposition process, which essentially constitutes a set of multiple adaptive Wiener filters, thereby showing excellent anti-noise ability. In modal separation, the variable modal decomposition technology can effectively distinguish two pure harmonic signals with similar frequencies.

[0096] Furthermore, for the process of optimizing variable mode decomposition by particle swarm algorithm, in this embodiment, the specific implementation method is as follows:

[0097] Initialize the potential solution of the variable mode decomposition into multiple particles; calculate the historical optimal position of any of the particles, and record the historical optimal positions of all the particles; determine the individual historical optimal fitness value of any of the particles and the group historical optimal fitness value of all the particles based on the historical optimal positions of all the particles; update the position and speed of all the particles based on the individual historical optimal fitness value and the group historical optimal fitness value; repeat the steps of determining the individual historical optimal fitness value of any of the particles and the group historical optimal fitness value of all the particles until the maximum number of iterations is reached, and obtain the variable mode decomposition optimized by the particle swarm algorithm.

[0098] Among them, since the modal decomposition method is very sensitive to the level of noise, when the signal noise level is high, the decomposition result may be interfered by the noise, the spectral band of the long-term mode will change dramatically over time, and will overlap globally, with boundary effects and burst signals. This is closely related to the use of the L2 smoothing stage, which over-punishes the domain boundary and internal jump mode decomposition methods when processing long time series. The problem requires a lot of computing resources. Therefore, this embodiment provides a solution based on the particle swarm algorithm to optimize the variable mode decomposition.

[0099] Moreover, for the classic decomposition algorithm, appropriate hyperparameters can improve the efficiency of modal decomposition. Usually, hyperparameters are set based on experience and comparison. In order to simplify the tedious manual parameter adjustment process, algorithms such as genetic algorithms, particle swarm optimization algorithms and gray wolf optimization algorithms are often used to find the optimal hyperparameters of deep learning models. Optimization algorithms have been widely used in related disciplines such as signal processing, deep learning and machine learning, and particle swarm optimization (PSO) is widely used due to its fast convergence speed. Its steps are mainly divided into initializing particle swarm parameters, finding the optimal solution of the particle swarm, updating the speed and position of each particle, and outputting the optimal solution.

[0100] However, the optimization algorithm may converge to a local optimal area too early in the search process, causing the search to stagnate. For functions with multiple local extreme points, the optimization algorithm is prone to fall into the local extreme points and cannot find the global optimal solution; the optimization algorithm only uses the information of group optimality and individual optimality in each iteration, but does not make full use of other information obtained in the calculation process. Therefore, when the problem has high requirements on the accuracy of the solution, the optimization algorithm may not be able to meet the requirements; the performance of the optimization algorithm is greatly affected by the selection and setting of parameters, especially the selection of inertia weight and acceleration factor often needs to be adjusted through trial and error, which increases the difficulty of algorithm implementation.

[0101] Therefore, in the particle swarm algorithm optimization variable mode decomposition solution provided in this embodiment, the particle swarm algorithm is first initialized, and a group of particles are randomly initialized in the solution space of the problem. Each particle represents a potential solution to the problem. Each particle has a position (indicating a solution) and a speed (indicating a moving direction and distance). Assume that N In the dimensional search space, there are M particles, each particle represents a solution.

[0102] No. i The particle positions are:

[0103] ,

[0104] No. i The particle speed is:

[0105] ,

[0106] Secondly, individual and social learning is carried out to determine the i The best historical position of a particle:

[0107] ,

[0108] At the same time, record and update the best position of all particles in the group:

[0109] ,

[0110] Again, the position of each particle is evaluated and its fitness is calculated according to the objective function of the optimization problem. i The individual historical optimal fitness value searched by the particle is , the historical optimal fitness value of the group is .

[0111] Then the position and velocity are updated, and the velocity of the particle is updated according to its current velocity, the distance from the individual optimal solution, and the distance from the global optimal solution.

[0112] The speed is actually the distance and direction of the particle's next iteration, that is, a position vector. The speed update formula, which consists of the inertial part, the cognitive part, and the social part, is as follows:

[0113] ,

[0114] in, , Respectively represent i The particle in k +1, k In the iteration n dimensional velocity vector; is the inertia weight; , Represent individual and group learning factors respectively; , Representation interval A random number; k Indicates the number of iterations; Indicates k After the iteration, i The optimal solution obtained by particle search; Indicates k After iterations, the entire particle swarm searches for the optimal solution; Indicates i The particle in k In the iteration n -dimensional position vector.

[0115] The inertia weight formula is as follows:

[0116] ,

[0117] in, is the maximum inertia weight; is the maximum inertia weight; is the current iteration number; is the maximum number of iterations. The position update formula is as follows:

[0118] ,

[0119] in, Indicates i The particle in k +1 iteration n Finally, the process of evaluating, updating the velocity and position is repeated until the maximum number of iterations is reached.

[0120] According to the above method, the process of particle swarm optimization of variable mode decomposition is as follows: Figure 2 The framework diagram of a power load forecasting method for extreme weather events is shown in the following figure, and the variable mode decomposition parameters to be optimized are determined: the modal number K , penalty factor α ; Then set the PSO parameters: number of particles, number of iterations, and cognitive coefficient c 1 and social coefficient c 2 , inertia weight ω ; Each particle represents a set of possible VMD parameter solutions, and the particle's position and velocity are randomly initialized in the parameter space; Define a fitness function to evaluate the performance of each particle; Calculate the fitness of each particle, update the individual best, and select the overall best from the individual best of all particles; Based on c 1 , c 2 and ω , update the speed and position of each particle, and judge whether the termination condition is reached, that is, whether the maximum number of iterations is reached. If so, the particle swarm algorithm optimization ends, otherwise, it continues to iterate.

[0121] Furthermore, the high-frequency component is processed by the pre-trained first model to obtain a first sequence, including:

[0122] The high-frequency components are feature extracted through the temporal convolutional network included in the first model, and the feature extraction results are fused with the features corresponding to the meteorological forecast data to obtain input features; along the channel dimension of the first model, the input features are divided into multiple subsequences through the frequency domain attention mechanism, and the frequency channel vector is determined through the multiple subsequences; according to the gated pulse neural P model included in the first model, the frequency channel vector is processed to obtain the first sequence.

[0123] Among them, the principle of deep learning is to learn the feature representation and pattern recognition of data by building a deep neural network model. It is a machine learning method that can automatically learn data features. By learning a large amount of sample data, the deep neural network model can automatically extract the main features of the data and realize the prediction task of new data. The core of the deep learning model is the neural network, which is composed of multiple interconnected neurons, each of which receives input signals and generates output signals. The deep learning model usually includes an input layer, a hidden layer, and an output layer. During the training process, the deep learning model optimizes the network parameters through the back propagation algorithm, so that the error between the model's predicted results and the actual results continues to decrease. By comparing the error between the network's output and the actual label, the error signal is transmitted backwards and the network parameters are updated to minimize the loss function.

[0124] However, existing deep learning models suffer from overfitting when predicting the impact of extreme weather on power load, that is, they perform well on training data but poorly on test data; they also have poor generalization capabilities and unstable performance on different data sets; and a large amount of data is required to train the prediction model, and data sparsity and poor data quality have a great impact on the performance of the algorithm.

[0125] Therefore, if Figure 2 As shown in the framework diagram of a power load forecasting method for extreme weather events, for the variable mode decomposition optimized by the particle swarm algorithm, the normalized load data sequence is processed to obtain high-frequency components, low-frequency components and residual components. The high-frequency components are input into the frequency-domain enhanced channel attention mechanism-based gated pulse neural P (Frequency Enhanced Channel Attention MechanismTime Convolutional Network- Frequency Enhanced Channel Attention MechanismGated Spiking Neural P, FAT-FAGP) model, that is, the first model, wherein the time convolutional network (Time Convolutional Network, TCN) effectively extracts the long-term time relationship and hidden information in the features, and then reduces the feature dimension of the high-frequency components. Combined with the meteorological forecast data, the frequency-domain enhanced channel attention mechanism module enhances the diversity of the extracted features. By continuously iterating many neurons, the FAGP model is used to effectively predict each high-frequency component to obtain a prediction sequence 1, that is, the first sequence.

[0126] Specifically, the FAT-FAGP model, i.e., the first model predicts one high-frequency component of the load each time, and repeats this process until all high-frequency prediction components are obtained, such as Figure 3 The FAT-FAGP model schematic diagram of a power load forecasting method for extreme weather events is shown in the figure. Based on the extracted feature information, the FAT-FAGP model can complete the extreme weather load forecasting task and the seasonal load forecasting task. By adding the frequency domain enhanced channel attention mechanism to the temporal convolutional network, the temporal convolutional network model can fully extract the volatility in the features by combining the feature extraction capability of the convolutional neural network and the time domain modeling capability of the recurrent neural network. In addition, the gated pulse neural P system combined with the frequency domain enhanced channel attention mechanism can effectively process small-scale and large-scale load data.

[0127] The prediction process of the FAT-FAGP model first extracts and fuses the high-frequency load components obtained through decomposition based on the feature extraction and feature fusion of the temporal convolutional network. X Input to TCN D filter to obtain , and then concatenate the last element of each vector to obtain the feature P ,feature P As shown below:

[0128] ,

[0129] Then the features P Combined with meteorological features to obtain a new input feature map J . Then, the features are separated in the frequency domain. Along the channel dimension, the frequency domain attention mechanism divides the input feature J into r Subsequence ,in; r Indicates the number of features; the number of channels is 1; X Represents the length of the time series. Then, each channel will be processed by the same frequency component, resulting in:

[0130] ,

[0131] in, and ; Represents at time Discrete cosine transform of ; Represents the length of the time series; yes P dimensional vector. Through stacking operation, the entire frequency channel vector can be obtained.

[0132] After that, we obtain the vector processed by the frequency domain enhanced attention mechanism Q Specifically, by continuously updating the attention weights, the FECAM tensor can be expressed as follows:

[0133] ,

[0134] in, and Represent sigmoid and ReLU activation functions respectively; and is the attention weight.

[0135] Redefine the FAGP model. As a nonlinear extension of the pulse neural P model, the gated pulse neural P model can effectively model the characteristics of long-term time series and handle non-stationary features. The FAGP is defined as follows:

[0136] ,

[0137] in, Represents a single alphabet; Represents a pulse, m The gated neurons are represented as ,and , No. m The internal state of a neuron is expressed as ; represents a series of nonlinear impulse rules; Expressed in its form, the pulse threshold is expressed as T ; and It is a nonlinear function; ; Forget Gate and consumption gate , and , and are input neurons and output neurons respectively.

[0138] The FAGP model can be informally described as a directed graph consisting of nodes and edges, where each node represents a gated neuron and each edge represents a connection between gated neurons. Each gated neuron has a data unit to store its internal state. Can be represented as a gated neuron In time t These gating neurons obey specific nonlinear spiking rules. , which contains linear or nonlinear functions and .when and When all are linear functions, the behavior of the FAGP model is similar to that of the traditional spiking neural P system. Therefore, the FAGP model can be regarded as a nonlinear extension of the spiking neural P system. Usually with are different, but they may also be the same function. When it is 0, it is consistent with the traditional model, and the impulse rule is the forgetting rule. In the above equation, the pulse condition can be expressed as , where Gating neurons At the moment t -1 state. The meaning of this rule can be expressed as follows: If at time t -1, Neuron The status reaches or exceeds a certain threshold T ,Right now , then the neuron Will be triggered or triggered. This triggering process consumes pulses and generate These newly generated pulses are then passed to subsequent neurons. In this process, the function g describes the consumption of pulses, while the function f Describe the generation of pulses. The process can be described by the following state equation:

[0139] ,

[0140] Finally, based on the high-frequency component prediction of the FAGP model, when considering the multivariate feature input divided into z subgroups , and z represents the length of the time series, the state equation and input-output relationship of the FAGP model are as follows:

[0141] ,

[0142] in, Represents the gated pulse neural P system in time t The predicted value of is the weight matrix; and is the bias vector; is the dot product; Representative i variables in time t The predicted value of Represents the FAGP model at time t The predicted value of and They represent the consumption gate and the reset gate respectively.

[0143] Furthermore, the process of processing the low-frequency component by the pre-trained second model to obtain the second sequence is specifically implemented as follows in this embodiment:

[0144] The second model based on the iTransfomer model architecture includes a normalization layer, a feedforward network and a self-attention mechanism to process the low-frequency component to obtain the second sequence.

[0145] Among them, Figure 2 As shown in the framework schematic diagram of a method for power load forecasting of extreme weather events, the low-frequency component is input into the iTransformer model to obtain a forecast sequence 2, i.e., a second sequence.

[0146] Based on this, iTransformer restructured the architecture without modifying the basic components and simply used the feedforward network in an inverted form and dimension. Each variable token is represented nonlinearly using a feedforward network. In addition, the attention mechanism uses variable tags embedded by time points of a single sequence to capture and obtain multivariate correlations. The iTransformer model in this embodiment only considers the encoder structure of the Transformer, such as Figure 4 An iTransformer model schematic diagram of an extreme weather event power load forecasting method is provided.

[0147] Specifically, in iTransformer, n is the index, represents the entire time series of each variable, each variable The prediction process can be simply expressed as follows:

[0148] .

[0149] For the normalization layer of the second model, in order to effectively deal with non-stationary problems, the series representation of the single variable is normalized in iTransformer to reduce the interaction noise between the delayed process and the non-causal process. , the time step is Q , the number of variables is W , predicting the future E The layer normalized equation for the low-frequency load component of time steps is as follows:

[0150] ,

[0151] in, include O indivual P Dimension embedding tokens and layer indexes are indicated by superscripts. , then Projection: Embedding and Projection are implemented through multi-layer perceptrons. In each TrmBlock, variable tokens interact with each other through a shared feedforward network and self-attention mechanism.

[0152] For the feedforward network of the second model, in iTransformer, a feedforward network is used to represent the time series in each variable token. Based on the universal approximation theorem, iTransformer extracts complex representations to represent time series. By stacking reversed Transformer blocks, iTransformer encodes the observed sequence and decodes the representation of future sequences using dense nonlinear connections.

[0153] Specifically, for each extracted sequence representation, , the self-attention mechanism module uses linear projection to obtain keys, pairs and values A, where the projection dimension is d k express.

[0154] Furthermore, the residual component is processed by the pre-trained third model, and the processing result is combined with the first sequence and the second sequence to determine the process of power load forecast data. In this embodiment, the specific implementation method is as follows:

[0155] According to the multi-head sparse self-attention mechanism, decoder and Informer model distillation mechanism included in the third model of the Informer model architecture, the residual component is processed to obtain the residual load component prediction data; the first sequence, the second sequence and the residual load component prediction data are combined to determine the power load prediction data.

[0156] Among them, in order to solve the problem of poor prediction performance of traditional deep learning models, such as Figure 2 A framework diagram of a power load forecasting method for extreme weather events is provided, in which the residual component is predicted by an Informer model.

[0157] Specifically, Figure 5 The Informer model schematic diagram of an extreme weather event power load forecasting method is shown in the figure, which adopts a multi-head sparse self-attention mechanism, a decoder, and an Informer model distillation mechanism.

[0158] Among them, for the multi-head sparse self-attention mechanism, the traditional self-attention mechanism is based on tuple input and performs a scaled dot product operation through query-key-value, and its expression is as follows:

[0159] ,

[0160] in, Represents the traditional self-attention mechanism operation method; , and Matrices representing queries, keys, and values ​​respectively; is the input dimension; Softmax( ) is the normalized exponential function, a The attention coefficient of a pair is calculated as follows:

[0161] ,

[0162] in, Represented in the matrix The vector of the middle pair; and Respectively represented in the matrix Middle b and c key vectors; Represents the first b A vector of values; Select the asymmetry index ; Indicates expected value; represents conditional probability; .

[0163] No. a The evaluation formula for the sparsity of pairs is defined as follows:

[0164] ,

[0165] Among them, the former is the Log-Sum-Exp over known keys, which calculates the arithmetic mean. Represents the number of key vectors on the matrix. Then, the calculation formula of the probabilistic sparse self-attention mechanism is:

[0166] ,

[0167] in, is a are sparse matrices of the same size, and Only include sparse evaluation based Before getting off u The most representative queries.

[0168] For the decoder and the Informer model distillation mechanism, the distillation mechanism can reduce the high spatial complexity and memory consumption caused by long sequence input. It assigns higher weights to the main features and generates feature maps in the next layer, shortening the length of the input sequence layer by layer.b Layer to b The distillation operation of the +1 layer is as follows:

[0169] ,

[0170] in, represents the attention module; Representatives in t Moment b The feature matrix of the layer; Convld represents a one-dimensional convolution operation; ELU is the activation function; MaxPool is the maximum pooling operation.

[0171] During the decoder processing, the cascade operation Input time series It contains two parts. One is the historical load series , and the other part is the prediction sequence that needs to mask the future characteristic factors , the specific operations are as follows:

[0172] ,

[0173] In addition, in order to prevent the interference of autoregression on the training effect, a masked sparse multi-head attention mechanism is adopted to avoid the influence of the unpredicted information on the current prediction point. The generative decoder can generate the entire prediction result at once during the training process, thereby improving the decoding efficiency.

[0174] like Figure 2 As shown in the framework diagram of a method for power load forecasting for extreme weather events, after the high-frequency components, low-frequency components and residual components are processed with corresponding models respectively, it is necessary to reconstruct all the obtained prediction components and combine them with inverse normalization to obtain the final prediction result, i.e., the power load forecast data.

[0175] Corresponding to the above method embodiment, the present application also provides an embodiment of a power load prediction device for extreme weather events, Figure 6 FIG. 1 is a schematic diagram showing the structure of an extreme weather event power load prediction device provided by an embodiment of the present application. Figure 6 As shown, the device comprises:

[0176] A collection module 602 is configured to collect power load data and weather forecast data;

[0177] The elimination module 604 is configured to determine the correlation between the power load data and the weather forecast data based on the correlation coefficient method, and eliminate redundant features to obtain a load data sequence;

[0178] A decomposition module 606 is configured to process the load data sequence through variable mode decomposition optimized by a particle swarm algorithm to obtain a high frequency component, a low frequency component and a residual component;

[0179] A first processing module 608 is configured to process the high frequency component through a pre-trained first model to obtain a first sequence, and process the low frequency component through a pre-trained second model to obtain a second sequence;

[0180] The second processing module 610 is configured to process the residual component through a pre-trained third model, and combine the processing result with the first sequence and the second sequence to determine the power load prediction data.

[0181] In an optional embodiment, the power load prediction device for extreme weather events further includes:

[0182] The preprocessing module is configured to fill the missing values ​​of the power load data based on linear interpolation and polynomial interpolation; and process the abnormal values ​​of the power load data based on the standard score and the Box plot method.

[0183] In an optional embodiment, the first processing module 608 is further configured to:

[0184] The high-frequency components are feature extracted through the temporal convolutional network included in the first model, and the feature extraction results are fused with the features corresponding to the meteorological forecast data to obtain input features; along the channel dimension of the first model, the input features are divided into multiple subsequences through the frequency domain attention mechanism, and the frequency channel vector is determined through the multiple subsequences; according to the gated pulse neural P model included in the first model, the frequency channel vector is processed to obtain the first sequence.

[0185] In an optional embodiment, the first processing module 608 is further configured to:

[0186] The second model based on the iTransfomer model architecture includes a normalization layer, a feedforward network and a self-attention mechanism to process the low-frequency component to obtain the second sequence.

[0187] In an optional embodiment, the second processing module 610 is further configured to:

[0188] According to the multi-head sparse self-attention mechanism, decoder and Informer model distillation mechanism included in the third model of the Informer model architecture, the residual component is processed to obtain the residual load component prediction data; the first sequence, the second sequence and the residual load component prediction data are combined to determine the power load prediction data.

[0189] In an optional embodiment, the decomposition module 606 is further configured to:

[0190] Based on Hilbert transform, the analytical signals of all modal functions are determined, and according to all the analytical signals, the single-sided spectra and the center frequencies of all the analytical signals are determined; according to the center frequency, the single-sided spectrum is modulated to the corresponding baseband to construct a constrained variable mode decomposition problem; based on the introduced quadratic penalty factor, the constrained variable mode decomposition problem is converted into an unconstrained variable mode decomposition problem to obtain an extended Lagrangian expression; the Lagrangian expression is gradually optimized through an alternating direction multiplier iteration algorithm to determine the variable mode decomposition.

[0191] In an optional embodiment, the decomposition module 606 is further configured to:

[0192] Initialize the potential solution of the variable mode decomposition into multiple particles; calculate the historical optimal position of any of the particles, and record the historical optimal positions of all the particles; determine the individual historical optimal fitness value of any of the particles and the group historical optimal fitness value of all the particles based on the historical optimal positions of all the particles; update the position and speed of all the particles based on the individual historical optimal fitness value and the group historical optimal fitness value; repeat the steps of determining the individual historical optimal fitness value of any of the particles and the group historical optimal fitness value of all the particles until the maximum number of iterations is reached, and obtain the variable mode decomposition optimized by the particle swarm algorithm.

[0193] The power load prediction device for extreme weather events provided in the present application collects power load data and meteorological forecast data; based on the correlation coefficient method, determines the correlation between the power load data and the meteorological forecast data, and eliminates redundant features to obtain a load data sequence; processes the load data sequence through variable mode decomposition optimized by a particle swarm algorithm to obtain high-frequency components, low-frequency components and residual components; processes the high-frequency components through a pre-trained first model to obtain a first sequence, and processes the low-frequency components through a pre-trained second model to obtain a second sequence; processes the residual components through a pre-trained third model, and combines the processing results with the first sequence and the second sequence to determine the power load prediction data. The variable mode decomposition optimized by the particle swarm algorithm can effectively select the number of decomposition modes, prevent insufficient or excessive decomposition, and effectively suppress high-frequency noise, thereby extracting clearer modes to effectively separate different frequency components, so that each mode has better physical meaning and interpretability, avoiding the problem of mode aliasing, and based on the FAT-FAGP load forecasting model, it can effectively extract features in the frequency domain and time domain, more comprehensively learn the temporal relationship between extreme weather characteristics and load characteristics, capture long-term dependencies with parallel computing, and is not prone to gradient explosion problems, thereby achieving better prediction performance in extreme weather scenarios. The integrated model achieves complementary advantages, and by combining the advantages of multiple models, the model's generalization ability in the face of sudden load changes under extreme weather is enhanced, effectively avoiding the problem of low prediction accuracy of a single prediction model under extreme weather, thereby improving the overall prediction accuracy and reducing the risk of overfitting and underfitting. Finally, on the premise of reducing feature redundancy and information calculation, it can achieve high-reliability signal decomposition, weaken data processing complexity, and complete complex extreme weather forecasting tasks. Further, by accurately predicting loads under extreme weather, it can effectively assist power grid operators in optimizing resource allocation, thereby ensuring the stable operation of the power grid.

[0194] In addition, since the proposed algorithm utilizes the extremely parallel characteristics of membrane computing, and the integrated model is also executed in parallel, the present invention can well meet the parallelism of different prediction algorithms based on the characteristics of parallel execution of the integrated model, and to a certain extent reduce computing resources at an efficient speed, thereby meeting the real-time requirements of the prediction model in the power company's dispatching platform; fully consider the impact of extreme meteorological factors on load changes, and weaken the mutation of load through decomposition algorithms, and then combine multiple algorithms to obtain prediction results, and at the same time use the results to exchange real-time information with the dispatching desk. This use of actual data for load dispatching and control has short-term observation and decision-making capabilities, ensuring real-time control of power dispatching. Regulation has high reliability; for load forecasting under extreme weather conditions, a parallel integrated forecasting algorithm is proposed in combination with membrane computing technology. The algorithm is superior to traditional forecasting algorithms in terms of predictability and generalization. Since the proposed algorithm utilizes the extremely parallel characteristics of membrane computing, and the big data dispatching center also executes the algorithm in parallel, it has strong applicability; in extremely hot / cold weather, the load, especially the air-conditioning load, fluctuates greatly, which poses a great challenge to the accuracy of traditional load forecasting models. The present invention uses existing historical data (historical load data and meteorological forecast data, etc.) and adopts an efficient forecasting method to estimate the power value in the future time period under cold waves or extreme heat waves, and the prediction accuracy (measured in R 2 As a measurement standard) as high as 95%.

[0195] The above is a schematic scheme of an electric load forecasting device for extreme weather events of this embodiment. It should be noted that the technical scheme of the electric load forecasting device for extreme weather events and the technical scheme of the electric load forecasting method for extreme weather events mentioned above belong to the same concept. For the details not described in detail in the technical scheme of the electric load forecasting device for extreme weather events, please refer to the description of the technical scheme of the electric load forecasting method for extreme weather events mentioned above. In addition, each component in the device embodiment should be understood as a functional module that must be established to implement each step of the program flow or each step of the method, and each functional module is not an actual functional division or separation definition. The device claim defined by such a group of functional modules should be understood as a functional module architecture that mainly implements the solution through the computer program recorded in the specification, and should not be understood as a physical device that mainly implements the solution through hardware.

[0196] Figure 7 The structure block diagram of a computing device 700 provided according to an embodiment of the present application is shown. The components of the computing device 700 include but are not limited to a memory 710 and a processor 720. The processor 720 is connected to the memory 710 via a bus 730, and the database 750 is used to store data.

[0197] The computing device 700 also includes an access device 740 that enables the computing device 700 to communicate via one or more networks 760. Examples of these networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 740 may include one or more of any type of network interface (e.g., a network interface card (NIC)) that is wired or wireless, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a World Wide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a universal serial bus (USB) interface, a cellular network interface, a Bluetooth interface, a near field communication (NFC) interface, and the like.

[0198] In one embodiment of the present application, the above components of the computing device 700 and Figure 7 Other components not shown in the figure may also be connected to each other, for example, via a bus. It should be understood that Figure 7 The computing device structure block diagram shown is only for the purpose of illustration, and is not intended to limit the scope of the present application. Those skilled in the art may add or replace other components as needed.

[0199] The computing device 700 may be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smart phone), a wearable computing device (e.g., a smart watch, smart glasses, etc.), or other types of mobile devices, or a stationary computing device such as a desktop computer or PC. The computing device 700 may also be a mobile or stationary server.

[0200] The processor 720 is used to execute computer executable instructions of each step of the method for predicting power load for extreme weather events.

[0201] The above is a schematic scheme of a computing device of this embodiment. It should be noted that the technical scheme of the computing device and the technical scheme of the power load forecasting method for extreme weather events described above are of the same concept, and the details not described in detail in the technical scheme of the computing device can be found in the description of the technical scheme of the power load forecasting method for extreme weather events described above.

[0202] An embodiment of the present application further provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, are used to execute the steps of the method for predicting power load for extreme weather events.

[0203] The above is a schematic scheme of a computer-readable storage medium of this embodiment. It should be noted that the technical scheme of the storage medium and the technical scheme of the power load forecasting method for extreme weather events described above belong to the same concept, and the details not described in detail in the technical scheme of the storage medium can be found in the description of the technical scheme of the power load forecasting method for extreme weather events described above.

[0204] An embodiment of the present application further provides a chip storing a computer program, which, when executed by the chip, implements the steps of the method for predicting power load for extreme weather events.

[0205] The above describes specific embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0206] The computer instructions include computer program codes, which may be in source code form, object code form, executable files or some intermediate forms, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0207] It should be noted that, for the above-mentioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the described action sequence, because according to the present application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0208] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0209] The preferred embodiments of the present application disclosed above are only used to help explain the present application. The optional embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of the present application. The present application selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present application, so that those skilled in the art can understand and use the present application well. The present application is only limited by the claims and their full scope and equivalents.

Claims

1. A method for predicting power load in extreme weather events, characterized in that: include: Collect power load data and weather forecast data; Based on the correlation coefficient method, the correlation between the power load data and the weather forecast data is determined, and redundant features are eliminated to obtain a load data sequence. Specifically, the Kendall correlation coefficient method is used to screen out the most relevant variable combination and remove redundant features; Processing the load data sequence by variable mode decomposition optimized by particle swarm algorithm to obtain high-frequency components, low-frequency components and residual components; The high-frequency component is processed by a pre-trained first model to obtain a first sequence, and the low-frequency component is processed by a pre-trained second model to obtain a second sequence. Specifically, the high-frequency component is feature extracted by a temporal convolutional network included in the first model, and the feature extraction result is feature fused with the feature corresponding to the meteorological forecast data to obtain input features. Along the channel dimension of the first model, the input features are divided into multiple subsequences by a frequency domain attention mechanism, and a frequency channel vector is determined by multiple subsequences. The frequency channel vector is processed according to the gated pulse neural P model included in the first model to obtain the first sequence. The normalization layer, feedforward network and self-attention mechanism included in the second model based on the iTransfomer model architecture are used to process the low-frequency component to obtain the second sequence. The process of obtaining the first sequence according to the high-frequency component is to input the high-frequency component into a temporal convolutional network of a frequency domain enhanced channel attention mechanism - a gated pulse neural P model based on a frequency domain attention mechanism, that is, in the first model, firstly, based on the feature extraction and feature fusion of the temporal convolutional network, the high-frequency load component obtained by decomposition is X Input to TCN D filter to obtain , and then concatenate the last element of each vector to obtain the feature , and then the feature P Combined with meteorological features to obtain a new input feature map J , then separate the features in the frequency domain, along the channel dimension, the frequency domain attention mechanism will input the features J Divide into r Subsequence , r Indicates the number of features, the number of channels is 1, X represents the length of the time series, and then each channel will be processed by the same frequency component to obtain , and , Represents at time The discrete cosine transform of represents the length of the time series, yes P dimensional vector, and then through the stacking operation, the entire frequency channel vector can be obtained, and the vector processed by the frequency domain enhanced attention mechanism can be obtained. Q， Multivariate features divided into z subgroups As the input of FAGP model, FAGP is defined as ， Represents a single alphabet; Represents a pulse, m The gated neurons are represented as ,and , No. m The internal state of a neuron is expressed as ; represents a series of nonlinear impulse rules; Expressed in its form, the pulse threshold is expressed as T ; and It is a nonlinear function; ; Forget Gate and consumption gate , and , and They are input neurons and output neurons respectively. Finally, based on the high-frequency component prediction of the FAGP model, the output result is the first sequence; The residual component is processed by a pre-trained third model, and the processing result is combined with the first sequence and the second sequence to determine the power load forecast data. Specifically, according to the multi-head sparse self-attention mechanism, decoder and Informer model distillation mechanism included in the third model of the Informer model architecture, the residual component is processed to obtain the residual load component prediction data, and the first sequence, the second sequence and the residual load component prediction data are combined to determine the power load forecast data.

2. The method according to claim 1, characterized in that After collecting the power load data and the weather forecast data, the method further includes: Filling missing values ​​of the power load data based on linear interpolation and polynomial interpolation; Based on the standard score and Hakone plot method, outliers of the power load data are processed.

3. The method according to claim 1, characterized in that The determination process of the variable mode decomposition is: Based on Hilbert transform, determining analytical signals of all modal functions, and according to all the analytical signals, determining the single-sided spectra and center frequencies of all the analytical signals; According to the center frequency, the single-sided spectrum is modulated to a corresponding baseband, and a constrained variable mode decomposition problem is constructed; Based on the introduced quadratic penalty factor, the constrained variable mode decomposition problem is transformed into an unconstrained variable mode decomposition problem, and an extended Lagrangian expression is obtained; The Lagrangian expression is gradually optimized by an alternating direction multiplier iteration algorithm to determine the variable mode decomposition.

4. The method according to claim 1, characterized in that The process of optimizing the variable mode decomposition by the particle swarm algorithm includes: Initializing a potential solution of the variable mode decomposition into a plurality of particles; Calculating the historical optimal position of any of the particles, and recording the historical optimal positions of all the particles; Based on the historical optimal positions of all the particles, determining the individual historical optimal fitness value of any particle and the group historical optimal fitness value of all the particles; Based on the individual historical optimal fitness value and the group historical optimal fitness value, updating the positions and velocities of all the particles; Repeat the steps of determining the individual historical optimal fitness value of any of the particles and the group historical optimal fitness value of all the particles until the maximum number of iterations is reached, and obtain the variable mode decomposition optimized by the particle swarm algorithm.

5. A power load forecasting device for extreme weather events, characterized in that: include: A collection module configured to collect power load data and weather forecast data; The elimination module is configured to determine the correlation between the power load data and the weather forecast data based on the correlation coefficient method, and eliminate redundant features to obtain a load data sequence. Specifically, the Kendall correlation coefficient method is used to screen out the most relevant variable combination and remove redundant features; A decomposition module is configured to process the load data sequence through variable mode decomposition optimized by a particle swarm algorithm to obtain a high-frequency component, a low-frequency component and a residual component; The first processing module is configured to process the high-frequency component through a pre-trained first model to obtain a first sequence, and process the low-frequency component through a pre-trained second model to obtain a second sequence. Specifically, the high-frequency component is feature extracted through the temporal convolutional network included in the first model, and the feature extraction result is feature fused with the feature corresponding to the meteorological forecast data to obtain input features. Along the channel dimension of the first model, the input features are divided into multiple subsequences through a frequency domain attention mechanism, and a frequency channel vector is determined through multiple subsequences. The frequency channel vector is processed according to the gated pulse neural P model included in the first model to obtain the first sequence. The normalization layer, feedforward network and self-attention mechanism included in the second model based on the iTransfomer model architecture are used to process the low-frequency component to obtain the second sequence. The process of obtaining the first sequence according to the high-frequency component is to input the high-frequency component into the temporal convolutional network of the frequency domain enhanced channel attention mechanism-the gated pulse neural P model based on the frequency domain attention mechanism, that is, in the first model, firstly, based on the feature extraction and feature fusion of the temporal convolutional network, the high-frequency load component obtained by decomposition is X Input to TCN D filter to obtain , and then concatenate the last element of each vector to obtain the feature , and then the feature P Combined with meteorological features to obtain a new input feature map J , then separate the features in the frequency domain, along the channel dimension, the frequency domain attention mechanism will input the features J Divide into r Subsequence , r Indicates the number of features, the number of channels is 1, X represents the length of the time series, and then each channel will be processed by the same frequency component to obtain , and , Represents at time The discrete cosine transform of represents the length of the time series, yes P dimensional vector, and then through the stacking operation, the entire frequency channel vector can be obtained, and the vector processed by the frequency domain enhanced attention mechanism can be obtained. Q， Multivariate features divided into z subgroups As the input of FAGP model, FAGP is defined as ， Represents a single alphabet; Represents a pulse, m The gated neurons are represented as ,and , No. m The internal state of a neuron is expressed as ; represents a series of nonlinear impulse rules; Expressed in its form, the pulse threshold is expressed as T ; and It is a nonlinear function; ; Forget Gate and consumption gate , and , and They are input neurons and output neurons respectively. Finally, based on the high-frequency component prediction of the FAGP model, the output result is the first sequence; The second processing module is configured to process the residual component through a pre-trained third model, and combine the processing result with the first sequence and the second sequence to determine the power load forecast data. Specifically, according to the multi-head sparse self-attention mechanism, decoder and Informer model distillation mechanism included in the third model of the Informer model architecture, the residual component is processed to obtain the residual load component prediction data, and the first sequence, the second sequence and the residual load component prediction data are combined to determine the power load forecast data.

6. A computing device, characterized in that include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the steps of the method for predicting power load for extreme weather events as described in any one of claims 1 to 4.

7. A computer-readable storage medium storing computer instructions, characterized in that: When the instruction is executed by a processor, the steps of the method for predicting power load for extreme weather events described in any one of claims 1 to 4 are implemented.

Images