Short-term load turning point extraction and multi-segment prediction method based on multi-modal fusion

Abstract

The invention provides a multi-modal fusion short-term load turning point extraction and multi-segment prediction method, which comprises the following steps of: firstly, converting a power load sequence of a data set into a threshold-free recurrence plot, and optimizing image features by combining gray-scale morphological opening operation of rectangular structure elements; secondly, inputting the optimized threshold-free recurrence plot and the original load sequence into a multi-modal fusion model, and predicting a to-be-predicted daily turning point; then, K-Medoids clustering analysis is carried out on the power load data set, and N types of data sets are obtained; and finally, dividing a window according to the turning point predicted by the multi-modal model, inputting the window into a CNN-LSTM model for window prediction, and analyzing a prediction result. According to the method, the problem of large load curve inflection point prediction error is solved, and the prediction precision of the inflection point time period is effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power load prediction, and particularly relates to a multi-modal fusion short-term load turning point extraction and multi-segment prediction method. BACKGROUND

[0002] Power load prediction is a core link for realizing safe and stable operation of a new power system and efficient configuration of energy, and its precision directly affects the economy and reliability of power grid operation. With the power load time series presenting stronger nonlinear characteristics, mathematical statistics methods have been difficult to handle. Traditional artificial intelligence methods are good at processing single-mode data and have high dependence on data set quality and feature engineering. Therefore, more and more researchers focus on the fields of combination model, feature mining and modal decomposition.

[0003] In view of the limitation that traditional artificial intelligence methods can only process single-mode data, scholars are studying multi-modal methods integrating different data types. The multi-modal fusion method combines text, image, sound and other various perception data, enhances the robustness and generalization ability of the model through feature fusion, and has been widely applied in multiple fields.

[0004] In recent years, researchers have converted one-dimensional time series into two-dimensional images, presenting the dynamic evolution rules in time series in the form of spatial topological structure, and further mining the hidden deep features in time series. Literature [Guo W, Xu L, Zhao D, et al. A Wind Power Combination Forecasting Method Based on GASF Image Representation and UniFormer[J]. Journal of Marine Science and Engineering, 2024, 12(7).] realized wind power prediction by composing 2D Gram angle and field images of wind power sequence and related environmental characteristics. Literature [Hui Wang, Guizhi Wang. The prediction model for haze pollution based on stacking framework and feature extraction of time series images[J]. Science of The Total Environment, 2022, 839: 156003.] converted single variable time series data into images through recurrence plot, and then made haze pollution prediction according to the extracted image features. Literature [Xin Su, Ziguang Jia, Lei Zhou, et al. Enhanced structural-load forecasting: Fusion of image analysis and time series with physics-driven deep-learning models[J]. Journal of Sound and Vibration, 2025, 601: 118944.] converted time series into Gramian angle field image and two-dimensional sequence, and then fused image and sequence by using deep learning network to enhance input feature representation, which effectively improved the structural load identification accuracy. The above researches show the feasibility of using one-dimensional sequence to convert into image to mine time series features. However, the quality of input image directly affects the feature extraction effect of machine learning model, so image processing technology is widely used to optimize the original image.

[0005] In addition, more and more scholars have noticed that the prediction error of the inflection point, peak value and other mutation areas of the load curve is significantly higher than that of the continuous rising or falling area, so scholars have conducted a large number of extreme load researches. Literature [Peng Y G, Pan K D, Zhang D, et al. Multi-Section Short-Term Load Forecasting Based on Seasonal Load Adaptive Division and Important Point Segmentation[J]. Power System Technology, 2020, 44(02): 603-613] segmented the important points on the typical daily load curve after k-means clustering, realizing subsection forecasting. Literature [Yu, G. Z., Lu, L., Tang, et al. Ultra-Short-Term Wind Power Subsection Forecasting Method Based on Extreme Weather[J]. IEEE Transactions on Power Systems, 2023, 38(6): 5045-5056.] proposed a power time series trend discrimination method and inflection point detection method to accurately identify extreme weather periods. The above studies show that multi-section forecasting according to the trend of the load curve can improve the prediction accuracy of the load peak interval.

[0006] In recent years, the method of accurately identifying turning points has been widely applied in various fields. Literature [Shangzhe Li, Junran Wu, Xin Jiang, et al. Chart GCN: Learning chart information with a graph convolutional network for stock movement prediction[J]. Knowledge-Based Systems, 2022, 248: 108842.] extracts the key point sequence from the stock price sequence into an image, and then fully mines the chart information to predict the stock trend. Literature [Li, Y. Liu, X. Chen, J. Wu, et al. Forecasting Turning Points in Stock Price by Integrating Chart Similarity and Multipersistence[J]. IEEE Transactions on Knowledge and Data Engineering, 2024, 36(12): 8251-8266.] uses chart similarity to convert the price sequence into an image, and also introduces a multipersistence topology tool to accurately predict the stock turning point. Literature [Pouya Khodaee, Akbar Esfahanipour, Hassan Mehtari Taheri. Forecasting turning points in stock price by applying a novel hybrid CNN-LSTM-ResNet model fed by 2D segmented images[J]. Engineering Applications of Artificial Intelligence, 2022, 116: 105464.] takes the stock price and technical indicators as input, converts them into a two-dimensional image, and proposes a convolutional neural network-long short-term memory-residual network (CNN-LSTM-ResNet) model to accurately predict the turning point. The above researches in the field of stock price prediction show the feasibility of applying sequence-to-image to predict turning points. These researches provide diverse ideas for the identification of power load curve turning points, but there is still room for further improvement.

[0007] In summary, most of the prediction methods proposed in the current literature focus more on feature extraction, model improvement, and modal decomposition. However, existing research still has deficiencies in the following aspects: traditional methods rely only on a single data modality and cannot capture both the spatial topological correlation and the time-dependent relationship characteristics of the load sequence. For example, pure LSTM time series models can easily overlook the correlation between adjacent time periods, and pure CNN image models can have difficulty effectively learning long-term time-dependent relationships. Common sequence-to-image methods either over-smooth local details (e.g., GASF blurs load mutation points) or lose global dynamic characteristics (e.g., MTF only aggregates local state transitions), making it difficult to balance local detail capture and global feature preservation, resulting in insufficient representation of turning points by image features. Conventional image processing methods (such as Gaussian filtering) tend to suppress areas with dramatic pixel changes (image features corresponding to load turning points also have dramatic pixel changes), which can weaken or even eliminate subtle mutation structure information related to turning points, thereby affecting the accuracy of turning point prediction. Existing segmented prediction methods use fixed segmentation results of the clustering center curve (e.g., turning points based on K-means clustering centers as the fixed window division basis for similar days), which relies heavily on clustering results. However, real load curves are influenced by various factors, and their actual turning points may deviate from the turning points of the clustering center curve, which can lead to inaccurate segmentation and affect the prediction accuracy near the turning points. SUMMARY

[0008] To address the problem of large prediction error for load curve turning points and the inability of traditional methods to dynamically adjust the prediction window, the present application proposes a multi-modal fusion short-term load turning point extraction and multi-segmented prediction method, which constructs a K-CNN-LSTM model (D-W-CNN-LSTM) for predicting turning point division windows. This model controls the feature learning interval within the stable trend by accurately dividing the turning points and independently training in each window to avoid interference from global information. The use of this model can significantly reduce the average error of the window, thereby reducing the overall error of the sequence.

[0009] To achieve the above purpose, the technical scheme of the present application is as follows:

[0010] A multi-modal fusion short-term load turning point extraction and multi-segmented prediction method, the steps of which are as follows:

[0011] Step 1: Convert the power load sequence of the data set into a threshold-free recurrent graph, and optimize the image features using the gray-level morphological opening operation of the rectangular structural element.

[0012] Step two: input the optimized valve-free recurrence graph and the original load sequence into the multi-modal fusion model to predict the turning point of the day to be predicted;

[0013] Step three: perform K-Medoids clustering analysis on the power load data set to obtain N categories of data sets;

[0014] Step four: divide the window according to the turning point predicted by the multi-modal model and input it into the CNN-LSTM model for window prediction;

[0015] Step five: analyze the prediction results.

[0016] Preferably, the valve-free recurrence graph is a gray-scale graph generated by a distance matrix;

[0017] The gray-scale morphological opening operation is an opening operation of gray-scale erosion for removing false structures and gray-scale dilation for preserving structures, which realizes image optimization processing by traversing the neighborhood to obtain gray-scale extreme values:

[0018] The gray-scale erosion scans the image A pixel by pixel with the center point of the structure element B to obtain the minimum value of the pixel points in the covered area of B, and replaces the pixel value of the scanning point with this minimum value. The principle of gray-scale erosion of the structure element B at (x, y) on the image A is as follows:

[0019] (1) ;

[0020] The gray-scale dilation is the inverse operation of the gray-scale erosion, that is, taking the maximum value of the pixel points in the covered area of the structure element B. The principle of gray-scale dilation of the structure element B at (x, y) on the image A is as follows:

[0021] (2) ;

[0022] In the formula, s and t represent the coordinate offset of the structure element, which is used to define the range of the current pixel neighborhood.

[0023] Preferably, the multi-modal fusion model includes two branches, the first branch is to convert the time series into a valve-free recurrence graph, and the CNN is used to capture the topological features of the sequence in the phase space, and to retain the geometric structure information of the state transition;

[0024] The other branch is to directly process the original time series by LSTM to capture the dynamic dependence of the time series data;

[0025] Finally, a feature-level splicing fusion strategy is adopted to splice the spatial features extracted by CNN and the time sequence features extracted by LSTM in dimension, and then through a fully connected layer and a classifier to realize accurate prediction of the turning point.

[0026] Preferably, the CNN-LSTM model is obtained by connecting a three-layer one-dimensional CNN and two-layer LSTM in series; the CNN extracts local features and reduces dimension through convolution and pooling layers, and inputs the one-dimensional vector after the expansion layer into the LSTM to learn the long-time dependence of the sequence.

[0027] Preferably, the method for performing K-Medoids clustering analysis on the power load data set is:

[0028] The K-Medoids clustering algorithm based on cosine similarity is adopted to divide the load data sample set:

[0029] Let And The load data of the i-th day and the j-th day of a certain area, the calculation formula of the cosine similarity is as follows:

[0030] (3) ;

[0031] In the formula, S cos represents the similarity of the i-th day and the j-th day.

[0032] Preferably, the LSTM network is an improved recurrent neural network, which adds input gate, output gate and forget gate in the hidden layer, and increases the unit for storing memory; the basic principle of the LSTM structure unit is as follows:

[0033] When the input quantity is input to the LSTM network hidden layer, it is first processed by the input gate through nonlinear transformation, and then superimposed with the memory unit state processed by the forget gate to form a new memory unit state. Finally, the memory unit state is processed by a nonlinear function, and then multiplied by the current information state processed by a nonlinear function to obtain the output of the hidden layer; the calculation formula of the forget gate in the LSTM model is shown in formula (4), the calculation formula of the input gate is shown in formula (5) and formula (6), the calculation formula of the output gate is shown in formula (7), and the current time unit state value and the output value are shown in formula (8) and formula (9):

[0034] (4) ;

[0035] (5) ;

[0036] (6) ;

[0037] (7) ;

[0038] (8) ;

[0039] (9) ;

[0040] In the formula, f t i t and o t These represent the output values ​​of the forget gate, input gate, and output gate at time t, respectively; g t Let x be the candidate value of the input gate at time t; sigmoid and tanh are the sigmoid function and hyperbolic tangent function, respectively; x t Input value at time t; s t-1 s t , respectively, represent the output values ​​at time t-1 and time t; W is the weight coefficient matrix, where W=[W fx W fs W ix W is W gx W gs W ox W os The corresponding elements represent functions f. t i t g t and o t Enter x t and output s t-1 Weighting coefficients; b f b i b g b o These are the bias terms of the corresponding functions; c t Let t be the state value at time t.

[0041] Preferably, after converting the power load sequence of the dataset into a threshold-free recursive graph, the information entropy and edge strength index of the converted image are calculated.

[0042] The specific expression for information entropy is:

[0043] (14);

[0044] in, This represents the probability of i values ​​occurring;

[0045] The specific expression for edge strength is:

[0046] (15);

[0047] (16);

[0048] (17);

[0049] (18);

[0050] (19);

[0051] wherein I represents an image gray matrix; "*" represents a convolution operation, i.e. an element-wise product sum of a neighborhood centered at (x, y) and an operator; G x and G y respectively represent a horizontal gradient operator and a vertical gradient operator in a Sobel algorithm, used to calculate gradients in horizontal and vertical directions; E x (x, y) and E y (x, y) respectively represent binaryzation using only G x and only G y .

[0052] Preferably, the turning point dataset is extracted by using an important turning point algorithm, and the important turning point extraction method for the load curve time series is as follows:

[0053] A time series is a series of data points arranged in chronological order, usually collected at equal intervals, and each data point represents a numerical value at a certain time; let the load curve time series be S n n represents the load power at time n;

[0054] Definition wherein S αm represents the mth turning point, satisfying the following conditions:

[0055] (24);

[0056] (25);

[0057] (26);

[0058] (27);

[0059] wherein α m represents the angle change value at the point S m ; θ m1 and θ m2 respectively represent the angle values of S m and adjacent S m-1 and S m+1 ;

[0060] Definition of important turning point sequence wherein S iαk represents the kth important turning point; in the sequence composed of three consecutive turning points S αj-1 , S αj , S αj+1 , S αjCan it become a major turning point and S αj To S αj-1 S αj+1 The distance D between the line segments is related to the region of line segment; the larger D is, the larger the distance S is. αj The farther away from the line segment, the more obvious the turning trend, and the greater the possibility of it becoming an important turning point; the vertical distance is used for calculation, and the formula is as follows.

[0061] (28);

[0062] In the formula, v1, v2, and v2 represent three consecutive turning points S. αj-1 S αj S αj+1 Based on this formula, traverse the sequence of turning points and extract the important turning point sequence S. Iα .

[0063] The beneficial effects of this invention are:

[0064] 1) The method proposed in this invention, which combines grayscale morphological opening operation to optimize thresholdless recursive graphs, can effectively preserve the global statistical features and local dynamic details of time series more completely and retain the nonlinear features of time series. Its information entropy and edge strength are better than traditional image conversion methods, providing high-quality input for feature extraction.

[0065] 2) The multimodal fusion model proposed in this invention achieves accurate prediction of turning points by fusing the spatial topological features of the thresholdless recurrent graph extracted by CNN with the long-term temporal dependencies of the original sequence captured by LSTM. The prediction accuracy is higher than 95% in various scenarios, which is better than the accuracy of using cluster center turning points.

[0066] 3) The window partitioning method for predicting important inflection points based on a multimodal model proposed in this invention overcomes the limitation of fixed windows in preset inflection point partitioning, which cannot dynamically adapt to load changes. Comparative experiments show that the DW-CNN-LSTM model achieves the best prediction results in all scenarios, and the prediction accuracy from the window method is improved to varying degrees.

[0067] 4) The window forecasting method proposed in this invention provides a new approach to power load forecasting and has certain theoretical and application value. Attached Figure Description

[0068] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0069] Figure 1 (a) Recurrence plot, (b) No-threshold recurrence plot.

[0070] Figure 2 (a) Morphological operation schematic diagram.

[0071] Figure 3 (a) LSTM structure unit schematic diagram.

[0072] Figure 4 (a) Multimodal hybrid model structure diagram.

[0073] Figure 5 (a) System overall flowchart of the present application.

[0074] Figure 6 (a) CNN-LSTM network structure diagram.

[0075] Figure 7 (a) To-be-tested day screening standard.

[0076] Figure 8 (a) Data processing flowchart.

[0077] Figure 9 (a) Comparison of results of various sequence-to-image methods.

[0078] Figure 10 (a) Information entropy change, (b) Edge intensity change.

[0079] Figure 11 (a) Gray-scale opening operation using different structure elements.

[0080] Figure 12 (a) Contour coefficient, (b) Working day type, (c) High-temperature day type, (d) Holiday type.

[0081] Figure 13 (a) Holiday data set outliers.

[0082] Figure 14 (a) Turning point marking.

[0083] Figure 15 (a) Training set input-output schematic diagram.

[0084] Figure 16 (a) Turning point prediction result, (b) High-temperature day type, (c) Holiday type.

[0085] Figure 17 (a) High-temperature day confusion matrix.

[0086] Figure 18 (a) Overall prediction result, (b) High-temperature day type, (c) Holiday type.

[0087] Figure 19 Absolute error, (a) weekday type, (b) high temperature day type, (c) holiday type. DETAILED DESCRIPTION

[0088] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0089] The embodiment of the present application proposes a multi-modal fusion short-term load turning point extraction and multi-segment prediction method, constructs a multi-modal fusion model, extracts the spatial topological features of the recurrence plot through CNN, captures the time sequence dependent features of the sequence through LSTM, realizes feature complementation through feature level splicing, and improves the turning point representation capability. The load time sequence is converted into a valve-free recurrence plot, which completely retains the nonlinear dynamic characteristics of the load sequence. Discard the conventional image processing idea, adopt the opening operation of "first corrosion pseudo, then inflation structure", remove the noise at the same time, completely retain the dramatic pixel change area corresponding to the turning point, and ensure that the mutation structure information is not lost. The turning point is divided based on the multi-modal model prediction, rather than relying on the fixed turning point of the historical clustering center. First, the load curve type is distinguished through K-Medoids clustering, and then the specific turning point of each to-be-predicted day is predicted in real time through the multi-modal model, realizing the dynamic adaptation of "one type and multiple windows", avoiding the misalignment of window division and actual load fluctuation. As shown in Figure 5 , the specific steps are as follows:

[0090] Step 1: Convert the power load sequence of the data set into a valve-free recurrence plot, and optimize the image features in combination with the gray level morphological opening operation of the rectangular structure element;

[0091] The valve-free recurrence plot (Valve-Free Recurrence Plot, VFRP) is an improved method of the traditional recurrence plot (Recurrence Plot, RP), which does not need to rely on the experience to select the threshold, and does not make the binary judgment, but directly reflects the distance between the state points through the gray value or the color depth. The closer the distance, the higher the recurrence intensity. Unlike the traditional RP which is classified as "non-recurrence" when the distance between the state points is close to the threshold value ε, the VFRP generates a gray image through the distance matrix, and the gray value is inversely proportional to the distance, which retains the real distance information and the recurrence relationship level difference, and avoids the loss of key dynamic characteristics. The RP and VFRP effect comparison is shown in Figure 1 .

[0092] Mathematical morphology method is to get the more essential shape of the object by the interaction between the object and the structure element. In image processing, the quality of the image can be improved by using the morphological method. The traditional morphological method is generally applicable to binary image, while the VFRP is gray image. Therefore, the gray level morphological method is used to process the VFRP in this embodiment, to keep their basic shape features and remove irrelevant structures.

[0093] The gray level morphological method is applicable to continuous gray value, and the image optimization processing is realized by traversing the neighborhood to get the gray extremum. The gray erosion is to scan the image A by the center point of the structure element B pixel by pixel, to get the minimum value of the pixel points in the covered area of B, and to replace the pixel value of the scanning point with the minimum value. The principle of the gray erosion of the image A by the structure element B at (x, y) is as follows:

[0094] (1) ;

[0095] The gray dilation is the inverse operation of the gray erosion, that is, to take the maximum value of the pixel points in the covered area of the structure element B

[35] The principle of the gray dilation of the image A by the structure element B at (x, y) is as follows:

[0096] (2) ;

[0097] In the formula, s, t represent the coordinate offset of the structure element, which is used to define the range of the neighborhood of the current pixel.

[0098] The image processing by using the erosion or dilation alone may cause the loss of image information. Therefore, the open operation is used to process the gray image in this embodiment, which means the operation of erosion first and then dilation: the gray erosion can effectively clean up the discrete pseudo-recurrence points generated by the small fluctuations of the signal in the one-dimensional sequence transformation process. In the dilation stage, the key features such as the diagonal structure reflecting the long-term evolution law of the one-dimensional sequence and the distribution mode of the recurrence points are ensured to be completely retained by the recovery operation. The principle of the morphological operation is as shown in Figure 2 .

[0099] Step two: input the optimized threshold-free recurrence graph and the original load sequence into the multi-modal fusion model to predict the turning point of the day to be predicted;

[0100] The time series turning point prediction method based on multi-modal fusion realizes cross-modal information integration through a branch feature extraction architecture. On the one hand, the time series is converted into a threshold-free recurrent graph, and a CNN is used to capture the phase space topological features of the sequence, retaining the geometric structure information of state transition; on the other hand, the original time series is directly processed by an LSTM to capture the dynamic dependence relationship of the time series data. A feature-level splicing fusion strategy is adopted to splice the spatial features extracted by the CNN and the time sequence features extracted by the LSTM in dimension, and then a full connection layer and a classifier are used to realize accurate prediction of the turning point. The multi-modal hybrid model structure is as shown in Figure 4 .

[0101] Step three: K-Medoids clustering analysis is performed on the power load data set to obtain a data set of N categories;

[0102] In order to reflect the change characteristics of the load under different scenarios, a K-Medoids clustering algorithm based on cosine similarity is used to divide the load data sample set:

[0103] Let and be the load data of the i-th day and the j-th day of a certain area, then the calculation formula of the cosine similarity is as follows:

[0104] (3).

[0105] In the formula, S cos represents the similarity degree of the i-th day and the j-th day. The greater S cos , the more similar the load curve forms represented by the two sequences, and the more likely they belong to the same category. When S cos is 1, the two curves are completely coincident.

[0106] Step four: according to the turning point predicted by the multi-modal model, divide the window input into the CNN-LSTM model for window prediction;

[0107] The LSTM network is an improved recurrent neural network (RNN), which adds an input gate, an output gate and a forget gate in the hidden layer, and increases a unit for storing memory, overcoming the gradient disappearance and gradient explosion phenomena in RNN, so that it can learn the long-term dependence relationship of time series. Its structure unit is as shown in Figure 3 . The basic principle of the LSTM structure unit is as follows:

[0108] When the input quantity is input to the LSTM network hidden layer, it is first processed by the input gate through nonlinear transformation, then superimposed with the memory cell state processed by the forgetting gate to form a new memory cell state, and finally the memory cell state is processed by a nonlinear function and multiplied by the current information state processed by a nonlinear function to obtain the output of the hidden layer; the calculation formula of the forgetting gate in the LSTM model is shown in formula (4), the calculation formula of the input gate is shown in formula (5) and formula (6), the calculation formula of the output gate is shown in formula (7), and the current cell state value and the output value are shown in formula (8) and formula (9):

[0109] (4);

[0110] (5);

[0111] (6);

[0112] (7);

[0113] (8);

[0114] (9);

[0115] In the formula, f t , i t and o t are the output values of the forgetting gate, the input gate and the output gate at t time, respectively; g t is the candidate value of the input gate at t time; sigmoid and tanh are sigmoid function and hyperbolic tangent function, respectively; x t is the input value at t time; s t-1 and s t are the output values at t-1 time and t time, respectively; W is a weight coefficient matrix, wherein W=[W fx W fs W ix W is W gx W gs W ox W os ], and the corresponding elements represent the weight coefficients of the input x t and the output s t-1 in the functions f t , i t , g t and o t , respectively; b f , b i , b g and b o are the bias terms of the corresponding functions; ct is the state value at time t.

[0116] The CNN-LSTM hybrid neural network combines the feature extraction capability of CNN and the long-term memory capability of LSTM for time series data. CNN extracts local features and reduces dimensions through convolution and pooling layers, and converts to a one-dimensional vector through the unfolding layer to input the LSTM to learn the long-time dependence of the sequence.

[0117] The CNN-LSTM hybrid neural network of the embodiment is composed of three one-dimensional CNN layers and two LSTM layers, and the specific structure is as shown in Figure 6 .

[0118] Step five: analyze the prediction results.

[0119] The existing literature often uses three indicators, mean absolute percentage error (MAPE), mean absolute error (MAE), and root mean square error (RMSE), to evaluate the prediction accuracy of the load prediction model.

[0120] The calculation formulas of the respective evaluation accuracies are as follows:

[0121] (10)

[0122] (11)

[0123] (12)

[0124] In the formula, N is the total number of samples, and respectively represent the predicted value and the actual value of the load at time i.

[0125] In the specific examples, the data set and the model parameter settings are as follows:

[0126] In this example, the effectiveness of the proposed model is verified using the ISO-NE public data set from January 1, 2012 to December 31, 2013. The ISO-NE data set contains 24 hours of load and weather data. The total data points of the selected ISO-NE data set are 17545, and the data sampling frequency is 24 points per day. The load data before the day to be predicted is used as the training set for turning point prediction training, and the day to be predicted is used as the test set to verify the accuracy of the model. The first 80% of each type of day data set is used as the training set for model training, and the remaining 20% is used as the test set to verify the prediction accuracy of the model.

[0127] Common meteorological factors such as temperature, humidity, temperature-humidity index, and weather type play an important role in short-term load forecasting. Electricity load on weekends and holidays differs significantly from that on weekdays. Therefore, date type is also a key factor influencing short-term load forecasting.

[0128] For meteorological characteristics, dry-bulb temperature refers to the value read from a dry-bulb thermometer when the air is exposed to the elements but not directly exposed to sunlight, expressed in degrees Fahrenheit (°F). Dew point temperature refers to the temperature at which air reaches saturation under constant water vapor content and pressure. A higher dew point temperature indicates more water vapor in the air; therefore, this example uses dew point temperature to represent humidity, expressed in degrees Fahrenheit (°F). The Temperature and Humidity Index (THI) is a dimensionless index that comprehensively assesses human comfort based on temperature and humidity. It reflects human comfort more accurately than temperature or humidity alone. For ease of quantitative analysis, numbers 1 to 7 represent Monday to Sunday, with weekdays quantified as number 1 and holidays or weekends quantified as number 2. Numbers 1-12 represent the corresponding months, numbers 1-31 represent the corresponding dates, and hours represent the time series, expressed in hours (h). In summary, the features selected for prediction in this example are temperature, humidity, temperature and humidity index, holidays, month, day of the week, date, and hour, as shown in Table 1.

[0129] Table 1 Feature Input

[0130]

[0131] Cluster analysis can yield datasets for various day types. However, clustering based on cosine similarity may lead to some categories not matching reality. Therefore, it's necessary to filter the dataset and remove outliers from the day types. For the high-temperature day dataset, remove... Data other than that. For weekday datasets, remove... Data other than that. For holiday datasets, remove... Data other than that.

[0132] One typical day to be predicted is selected from the test set of each day type for visualization analysis and accuracy verification. The selection criteria are representativeness within the cluster and scenario differences, meaning the day to be predicted must meet the core characteristics of that day type. Specific selection criteria are as follows: Figure 7 As shown.

[0133] Holiday day to be predicted: The 23rd of October in the 2nd year (Holiday = 0, Week = 1, Monday) is selected from the holiday day test set, the daily average temperature humidity index THI = 45 (consistent with the holiday day category), the load curve is typical "double peak" (early peak, late peak), the load curve and the holiday day center cosine similarity is 0.92, representing the scene in the same day.

[0134] High temperature day to be predicted: The 4th of September in the 2nd year is selected from the high temperature day test set, the daily average temperature humidity index THI = 70.4 (consistent with the high temperature day category), the load curve is "single peak" (noon peak), the load curve and the high temperature day center cosine similarity is 0.89, representing the scene in the same day.

[0135] Holiday day to be predicted: The 23rd of October in the 2nd year (Holiday = 0, Week = 1, Monday) is selected from the holiday day test set, the daily average temperature humidity index THI = 45 (consistent with the holiday day category), the load curve is typical "double peak" (early peak, late peak), the load curve and the holiday day center cosine similarity is 0.92, representing the scene in the same day.

[0136] The multi-modal prediction model parameter settings are as follows: the filter size of the convolution layers 1, 2 and 3 is [3 3], and the filter number is 16, 32 and 64 respectively; the maximum pooling layer size is [3 3]; the network unit number of the LSTM1 and 2 is 64 and 128 respectively; the random inactivation probability of the random inactivation layer is 0.2. The neural network adopts the Adam optimization algorithm, the maximum iteration number is 100 times, the initial learning rate is 0.005, and the learning rate decay rate is 0.5. The training will be carried out on the CPU. The device of the training process is as follows: CPU: IntelCore i5-10200H; memory: 16GB; software: MATLAB R2023b.

[0137] The model parameter settings of the CNN-LSTM network are as follows: the filter size of the convolution layers 1, 2 and 3 is [3 3], and the filter number is 64, 32 and 16 respectively; the maximum pooling layer size is [2 1]; the network unit number of the LSTM1 and 2 is 100 and 50 respectively; the random inactivation probability of the random inactivation layer is 0.2. The neural network adopts the Adam optimization algorithm, the maximum iteration number is 300 times, the initial learning rate is 0.005, and the learning rate decay rate is 0.5. The training will be carried out on the CPU. The device of the training process is as follows: CPU: Intel Core i5-10200H; memory: 16GB; software: MATLAB R2023b.

[0138] During data acquisition, due to instrument malfunctions or signal transmission errors, data gaps and outliers are inevitable. To ensure the accuracy and effectiveness of the model, data processing is required to improve data quality. The data processing flowchart is as follows: Figure 8 As shown, the process mainly includes two parts: data cleaning and standardization. This example categorizes data cleaning as data preprocessing, which is completed before inputting the data into the system. The filtering of daily datasets by type is mainly used to filter and adjust the daily datasets after cluster analysis, removing data that does not conform to the specified day type, thus ensuring the purity of each type of daily dataset.

[0139] Data cleaning mainly includes methods such as imputing missing values, identifying outliers, and smoothing data fluctuations. This example uses linear interpolation to impute missing data and a sliding window method to detect and replace outliers in the imputed dataset.

[0140] The principle of linear interpolation is to construct a straight line between two known points, with missing values ​​falling on this line proportionally to the time interval. The formula for linear interpolation is shown below:

[0141] (10)

[0142] Here, (x0, y0) and (x1, y1) are two known points, and y represents any value of x within the interval.

[0143] The sliding window method for identifying outliers essentially defines the statistical parameters (mean, standard deviation) within the sliding window as the "normal range." Values ​​outside this range are considered outliers, marked as such, and then replaced using linear interpolation. Assume the window size is n. When the value to be detected is x... t At that time, the data in the window is {x} t-n x t-n+1 , ... , x t-1 The formula for the sliding window method is shown below:

[0144] (11)

[0145] (12)

[0146] in, This represents the recent average level. This represents the normal fluctuation range in the near term.

[0147] This example sets the normal range to: , k is the threshold coefficient, usually set to 3, indicating that approximately 99.7% of normal data will fall within the range. The sliding window size n=3 is chosen, which matches the minimum interval size of the inflection point, optimizing the data information near the inflection point as much as possible to facilitate subsequent model learning. Therefore, the corresponding detection point x t The value of t ranges from [2, z-1] (where z represents the sequence length for sliding window outlier detection).

[0148] After performing cluster analysis and image transformation using clean data, when using a neural network model for prediction, the original data needs to be normalized to avoid differences in the units of measurement between different data points. The calculation formula is shown below:

[0149] (13)

[0150] Where: x' is the normalized value; x is the original data; x max x is the maximum value in the sequence. min This is the minimum value in the sequence. The normalized values ​​are all in the interval [0,1].

[0151] To verify the effectiveness of threshold-free recurrence graphs, this example compares the performance of five transformation methods—Gramian Angular Summation Field (GASF), Gramian Angular Difference Field (GADF), Markov Transition Field (MTF), RP, and VFRP—on a public dataset. Figure 9 As shown.

[0152] To quantify the conversion effect, the information entropy and edge strength indices of the converted image were calculated. Information entropy measures the richness of information in an image and quantifies the degree to which a two-dimensional image retains the global statistical features of the original signal. Edges refer to regions in the image where pixel values ​​change drastically; edge strength indicates the salience of image edges and quantifies the ability of a two-dimensional image to capture local details of the original signal. A comparison of the conversion effects for each image is shown in Table 2.

[0153] The specific expression for information entropy is:

[0154] (14);

[0155] in, This represents the probability of i values ​​occurring;

[0156] The specific expression for edge strength is:

[0157] (15);

[0158] (16);

[0159] (17);

[0160] (18);

[0161] (19);

[0162] In the formula, I represents the image grayscale matrix; "*" represents the convolution operation, which is the sum of the element-wise product of the neighborhood centered at (x, y) and the operator; G x and G y These represent the horizontal and vertical gradient operators in the Sobel algorithm, used to calculate the gradients in the horizontal and vertical directions, respectively; E x (x, y) and E y (x, y) represent using only G x and using only G y Perform binarization.

[0163] Table 2 Comparison of Conversion Effects for Each Image

[0164]

[0165] The closer the normalized entropy value is to 1, the more uniform the pixel distribution. The greater the edge strength of the 2D image, the more completely the dynamic features of the time series are preserved during 2D transformation. The results show that the threshold-free recursive graph not only effectively captures local and global patterns of the time series, but also significantly outperforms other methods in terms of entropy and edge strength, proving that using the threshold-free recursive graph for transformation yields the best results.

[0166] Threshold-free recursive graphs are achieved through phase space reconstruction, the key to which lies in the selection of embedding dimension and time delay. This invention uses information entropy and edge intensity as image transformation indicators, comparing the indicators under different embedding dimensions and time delays. Experimental results are as follows: Figure 10 As shown.

[0167] Depend on Figure 10 As can be seen, information entropy and edge strength are negatively correlated with both embedding dimension and latency. To achieve better image conversion results, this example selects a latency of 1 and an embedding dimension of 1.

[0168] In the morphological opening operation of images, the key lies in the selection of the structuring element, as different structuring elements produce different processing effects. Therefore, this example uses rectangular, disk, line segment, and rhombus structures for experimental comparison, setting the structuring element radius to 1. The result of the opening operation on the grayscale image is as follows. Figure 11 As shown.

[0169] To quantify the transformation effect, the structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), gradient magnitude similarity deviation (GMSD), and mean absolute error (MAE) of the transformed images were calculated.

[0170] The calculation formulas for each evaluation indicator are as follows:

[0171] (20)

[0172] In the formula, μ x and μ y σ represents the average values ​​of image x and y within a local window, respectively, indicating the image's brightness level. x 2 and σ y 2 σ and y are the variances of the image x and y within the local window, respectively, representing the image contrast. xy C1 represents the covariance of images x and y within a local window, indicating the structural similarity of the images. C1 = (K1L) 2 C2 = (K2L) 2 Typically, K1=0.01, K2=0.03, and L is the dynamic range of pixel values, which is usually 255 for 8-bit images.

[0173] (twenty one)

[0174] (twenty two)

[0175] In the formula, x(i, j) and y(i, j) represent the pixel values ​​of image x and y at position (i, j), respectively, and M and N are the height and width of the image, respectively. MAX is the maximum possible pixel value in the image. For example, MAX is 255 for an 8-bit image.

[0176] (twenty three)

[0177] In the formula, and These are the gradient magnitudes at pixel (i, j) in the original image and the processed image, respectively. S This represents the average gradient magnitude of the original image.

[0178] The closer SSIM is to 1, the smaller MAE is, the larger PSNR is, and the closer GMSD is to 0, the better the image processing quality. The results are shown in Table 3.

[0179] Table 3. Experimental Results of Image Optimization

[0180]

[0181] Experiments show that using a rectangular structuring element yields the best results in image shape opening operations, preserving more image detail features. Therefore, this invention selects a rectangle as the structuring element to optimize image quality.

[0182] The K-Medoids algorithm was used to perform daily clustering analysis on the electricity load dataset, with the silhouette coefficient as the evaluation metric, as shown in the attached figure. Figure 12 As shown in (a), the maximum silhouette coefficient is 0.5112. When K=3, the silhouette coefficient drops sharply, which means that the correlation within the cluster drops sharply. Therefore, K=3 is the optimal number of clusters.

[0183] The clustering results with K=3 are shown in the appendix. Figure 12 As shown, analysis of the clustering results reveals that: the first type of daily load curve has the most days (364), mainly concentrated on weekdays; the second type has fewer days (181), mainly concentrated on high-temperature days; and the third type also has fewer days (182), mainly concentrated on holiday dates. Therefore, based on the above clustering results, the daily load curves are divided into three typical scenarios: weekdays, high-temperature days, and weekends / holidays.

[0184] Due to the irregularity of holidays, clustering results during holidays often show a high number of anomalies, such as... Figure 13 As shown. For example, although the load data for February 20, 2012, November 23, December 24, 2012, February 18, 2013, and other dates are similar to the load curves of cluster centers for holidays, these dates do not belong to holidays. This will affect the training of subsequent sample datasets and the training of the neural network, thus affecting the prediction accuracy. Therefore, this invention adopts... Figure 7 The standard filtered dataset shown yielded 364 weekdays, 181 high-temperature days, and 172 holiday days.

[0185] The clustering results primarily impact the prediction process through three stages: data reconstruction, independent training, and window initialization. First, after clustering and dataset selection, three independent datasets are obtained, one for each load type. The load type labels are then used as feature vectors to reconstruct the feature sets, enhancing their representational capabilities. Next, during model training, each dataset is processed separately. This allows the model to focus on learning the load characteristics of a specific scenario, reducing cross-interference between different load types. Finally, based on the inflection point prediction results, windows are divided for each load type, and independent window predictions are performed.

[0186] The method for extracting important inflection point datasets using the important inflection point algorithm is as follows:

[0187] A time series is a sequence of data points arranged chronologically, typically collected at equal intervals, with each data point representing a value at a specific time. A load curve time series is also included. S n Let be the load power at time n;

[0188] definition S αm Let the m-th turning point satisfy the following conditions:

[0189] (twenty four);

[0190] (25);

[0191] (26);

[0192] (27);

[0193] In the formula, α m Representing point S m The change in angle at that point; θ m1 and θ m2 S represents m With adjacent S m-1 and S m+1 Angle value;

[0194] Define the sequence of important turning points S iαk The k-th important turning point; three consecutive turning points S αj-1 S αj S αj+1 In the constructed sequence, S αj Can it become a major turning point and S αj To S αj-1 S αj+1The distance D between the line segments is related to the region of line segment; the larger D is, the larger the distance S is. αj The farther away from the line segment, the more obvious the turning trend, and the greater the possibility of it becoming an important turning point; the vertical distance is used for calculation, and the formula is as follows.

[0195] (28);

[0196] In the formula, v1, v2, and v2 represent three consecutive turning points S. αj-1 S αj S αj+1 Based on this formula, traverse the sequence of turning points and extract the important turning point sequence S. Iα .

[0197] like Figure 14 As shown, turning points are marked as 1, and other points are marked as 0.

[0198] Define a 72-hour inflection point marker sequence Z, and a time series T: ;in, The time indicates the date to be predicted.

[0199] Multimodal training sets do not require clustering to categorize the dataset by day; they directly process historical loads, effectively improving efficiency. This example applies a sliding window approach with one-intervals to the historical time series, resulting in a sliding window dataset. Point-by-point predictions are performed using the sliding window dataset from the two days prior to the prediction date, forecasting the 24-hour data for that day. Simply put, each 24-hour window predicts the third point onwards, and so on. This avoids the training set's sliding window containing actual values ​​(to perform 24-hour point-by-point predictions, a sliding window dataset from the previous two days is needed to avoid including the prediction date). For example, using... The data is used as the training set input for the model, including the data from the third day. As the training set output of the model. The data is used as the training set input for the model, including the data from the third day. The data serves as the training set output for the model. Similarly, a sliding window of data is used to predict the inflection point label for the 24 hours of the test day. The principle of training set input and output is as follows: Figure 15 As shown.

[0200] To verify the effectiveness of the proposed multimodal fusion model for predicting inflection points, this example sets up an ablation experiment, using CNN, LSTM, cluster center-based inflection point prediction, and a multimodal model for inflection point prediction as experimental groups. The prediction results are as follows: Figure 16 As shown.

[0201] like Figure 16It can be seen that the multimodal model outperforms cluster center partitioning, CNN prediction, and LSTM prediction in predicting inflection points. Table 4, showing the ablation experiments, indicates that relying solely on CNN or LSTM for inflection point prediction does not significantly improve performance compared to cluster center partitioning. This demonstrates the complexity of inflection point prediction; neither CNN's strength in extracting local spatial features nor LSTM's strength in capturing temporal dependencies can effectively identify the core dynamic features of inflection points when used alone. They are sensitive to stable data changes but lack the ability to capture non-stationary "turning point" signals triggered by multiple factors, thus their performance is similar to that of fixed-window clustering methods. The multimodal model proposed in this invention achieves the best results in all three scenarios, with an accuracy rate that keeps the error within one unit. This demonstrates the combined influence of spatial local anomalies and temporal trend abrupt changes on inflection points. This invention views the fusion process as a process of feature complementarity and decision calibration. The local features extracted by CNN provide the "attention focus" for LSTM, enabling it to pay more attention to key regions where turning points may occur when analyzing long sequences. Conversely, the temporal context provided by LSTM offers "reasonableness verification" for the local judgments of CNN, avoiding the misjudgment of isolated, non-trend noise as turning points. This accurately corrects the turning point localization drift problem caused by the fixed window in clustering methods. Furthermore, the method proposed in this invention can predict turning points without clustering, making it applicable to more datasets and more efficient.

[0202] This invention calculates the accuracy, precision, recall, and F1 score of several methods at the 24-hour inflection point as indicators, and the results are shown in Table 4. To intuitively demonstrate the classification prediction performance of each method, this example uses a high-temperature day as an example and calculates the confusion matrix for each method, where class 1 is labeled "0" and class 2 is labeled "2". The confusion matrix is ​​shown below. Figure 17 As shown.

[0203] Table 4 Indicators for Defining Turning Points

[0204]

[0205] Overall Prediction Results Analysis

[0206] To facilitate the analysis of the prediction performance of the prediction method proposed in this invention, the K-CNN-LSTM model that divides the window according to the cluster center is named Fixed-Window-CNN-LSTM (FW-CNN-LSTM), and the K-CNN-LSTM model that divides the window according to the prediction of the inflection point of the multimodal model is named Dynamic-Window-CNN-LSTM (DW-CNN-LSTM).

[0207] This invention tested six prediction models—CNN-LSTM, K-CNN, K-LSTM, K-CNN-LSTM, FW-CNN-LSTM, and DW-CNN-LSTM—using the ISO-NE dataset. The input to these models consisted of the feature input set for each typical day, and the output was the prediction load for that prediction day. The prediction results are as follows: Figure 18 As shown in Table 5, the prediction accuracy error results are as follows.

[0208] The K-CNN-LSTM, FW-CNN-LSTM, and DW-CNN-LSTM models with high prediction accuracy were selected for further comparative analysis. The absolute error histogram is shown below. Figure 19 As shown. By Figure 18 and Figure 19 As can be seen, the prediction results of the DW-CNN-LSTM model proposed in this invention are closest to the true values. A detailed analysis of Table 5 is as follows.

[0209] In the weekday category, the MAPE of the FW-CNN-LSTM and DW-CNN-LSTM models decreased by 30.67% and 37.34% respectively compared to the K-CNN-LSTM model. The MAPE of the FW-CNN-LSTM model decreased by 9.61% compared to the DW-CNN-LSTM model.

[0210] In the high-temperature day category, the MAPE of the FW-CNN-LSTM and DW-CNN-LSTM models decreased by 13.39% and 39.28% respectively compared to the K-CNN-LSTM model. The MAPE of the FW-CNN-LSTM model decreased by 29.89% compared to the DW-CNN-LSTM model.

[0211] In the holiday category, the MAPE of the FW-CNN-LSTM and DW-CNN-LSTM models decreased by 24.98% and 40.73% respectively compared to the K-CNN-LSTM model. The MAPE of the FW-CNN-LSTM model decreased by 20.99% compared to the DW-CNN-LSTM model.

[0212] Table 5 Prediction Accuracy

[0213]

[0214] The specific analysis shows that the prediction results of the multimodal prediction inflection point method proposed in this invention are significantly better than those of traditional load prediction and inflection point division based on cluster centers. This indicates that the method of dynamically dividing inflection points using a multimodal model has higher prediction accuracy than the method that relies on a preset inflection point division window.

[0215] Comparison with existing models

[0216] To fully demonstrate the advantages of the proposed model in short-term power load forecasting, four existing models were constructed: TCN-Transformer, LSTM-Transformer, CBAM-TCN-BIGRU, and BiLSTM. The key structural and training parameters of the comparison models were made as similar as possible to those of DW-CNN-LSTM, and the specific parameter settings are shown in Tables 6 and 7. Taking weekday types as an example, to compare the improvement effect of the proposed multimodal model on the inflection point interval, the error magnitude of the inflection interval was compared. The prediction errors are shown in Table 8.

[0217] Table 6 Key structural parameters of the comparative models

[0218]

[0219] Table 7 Key training parameters of the comparison models

[0220]

[0221] Table 8. Daily average error of existing models on weekdays

[0222]

[0223] As shown in Table 8, compared with existing models, the model proposed in this invention has higher prediction accuracy, and both the average error of the inflection point interval and the average error of the window are significantly higher than other models. The inflection point interval error and the average error of the window are not significantly different among the models, indicating that the feature learning capabilities of each model in abrupt change regions and stable intervals are not significantly different. The main reason for the high accuracy of each model lies in parameter settings, optimization, and feature processing input. When all models use the same standard parameters, existing methods fit the entire sequence with a "globally unified model," without distinguishing between "stable intervals" and "inflection intervals." When the sequence undergoes a sudden change (such as changing from rising to falling), the global model is "biased" by the historical stable trend, leading to prediction deviations near the inflection point. DW-CNN-LSTM, through precise inflection point segmentation, stably controls the feature learning interval within the stable trend, training independently within each window to avoid interference from global information. The average error of the window can be significantly reduced, thereby reducing the overall sequence error.

[0224] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting short-term load inflection points and predicting them in a multi-modal fusion manner, characterized in that, The steps are as follows: Step 1: Transform the power load sequence of the dataset into a threshold-free recursive graph, and optimize image features by combining gray-level morphological opening operations of rectangular structural elements. Step 2: Input the optimized threshold-free recursive graph and the original load sequence into the multimodal fusion model to predict the turning point of the day to be predicted; Step 3: Perform K-Medoids clustering analysis on the power load dataset to obtain N categories of datasets; Step 4: Divide the window based on the inflection point predicted by the multimodal model and input it into the CNN-LSTM model for window prediction; Step 5: Analyze the prediction results.

2. The multimodal fusion method for short-term load inflection point extraction and multi-segment prediction according to claim 1, characterized in that, The threshold-free recursive graph is a grayscale image generated by a distance matrix; The gray-level morphological opening operation involves first performing gray-level erosion to remove artifacts, followed by gray-level dilation to preserve structure. By traversing the neighborhood and taking the gray-level extrema, image optimization processing is achieved. Gray-scale erosion scans image A pixel by pixel using the center point of structuring element B to obtain the minimum pixel value in the area covered by B, and then replaces the pixel value of the scanned point with this minimum value. The principle of gray-scale erosion of image A by structuring element B at (x, y) is as follows: （1）； Gray-level dilation is the inverse operation of gray-level erosion, that is, taking the maximum pixel value of the region covered by structuring element B; the principle of gray-level dilation of image A by structuring element B at (x, y) is as follows: （2）； In the formula, s and t represent the coordinate offsets of the structuring element, which are used to define the range of the current pixel's neighborhood.

3. The multimodal fusion method for short-term load inflection point extraction and multi-segment prediction according to claim 1 or 2, characterized in that, The multimodal fusion model includes two branches. The first branch converts the time series into a thresholdless recursive graph and uses CNN to capture the phase space topological features of the sequence while preserving the geometric structure information of the state transition. Another branch is to directly process the original time series using LSTM to capture the dynamic dependencies of time series data; Finally, a feature-level concatenation and fusion strategy is adopted to concatenate the spatial features extracted by CNN with the temporal features extracted by LSTM, and then use a fully connected layer and a classifier to achieve accurate prediction of the turning point.

4. The multimodal fusion method for short-term load inflection point extraction and multi-segment prediction according to claim 3, characterized in that, The CNN-LSTM model is obtained by concatenating three one-dimensional CNN layers and two LSTM layers. The CNN extracts local features and reduces dimensionality through convolution and pooling layers, and then converts them into one-dimensional vectors through unfolding layers before inputting them into the LSTM to learn the long-term dependencies of the sequence.

5. The multimodal fusion method for short-term load inflection point extraction and multi-segment prediction according to claim 1, characterized in that, The method for performing K-Medoids clustering analysis on the power load dataset is as follows: The K-Medoids clustering algorithm based on cosine similarity was used to partition the load data sample set: set up and Given the load data for a certain region on day i and day j, the formula for calculating the cosine similarity is as follows: （3）； In the formula, S cos This indicates the degree of similarity between day i and day j.

6. The multimodal fusion method for short-term load inflection point extraction and multi-segment prediction according to claim 4, characterized in that, The LSTM network is an improved recurrent neural network that adds input gates, output gates, and forget gates inside the hidden layers, and also adds units for storing memories; the basic principle of the LSTM structure unit is as follows: When the input is input into the hidden layer of the LSTM network, it first passes through the input gate and undergoes a nonlinear transformation. Then, it is superimposed with the memory cell state processed by the forget gate to form a new memory cell state. Finally, the memory cell state is processed by a nonlinear function and multiplied by the current information state processed by the nonlinear function to obtain the output of the hidden layer. The calculation formula of the forget gate in the LSTM model is shown in Equation (4), the calculation formula of the input gate is shown in Equations (5) and (6), the calculation formula of the output gate is shown in Equation (7), and the current cell state value and output value are shown in Equations (8) and (9). （4）； （5）； （6）； （7）； （8）； （9）； In the formula, f t i t and o t These represent the output values ​​of the forget gate, input gate, and output gate at time t, respectively; g t Let x be the candidate value of the input gate at time t; sigmoid and tanh are the sigmoid function and hyperbolic tangent function, respectively; x t Input value at time t; s t-1 s t These are the output values ​​at time t-1 and time t, respectively. W is the weight coefficient matrix, where W = [W fx W fs W ix W is W gx W gs W ox W os The corresponding elements represent functions f. t i t g t and o t Enter x t and output s t-1 weighting coefficients; b f b i b g b o These are the bias terms of the corresponding functions; c t Let t be the state value at time t.

7. The multimodal fusion method for short-term load inflection point extraction and multi-segment prediction according to claim 2, characterized in that, After converting the power load sequence of the dataset into a threshold-free recursive graph, the information entropy and edge strength index of the converted image are calculated. The specific expression for information entropy is: （14）； in, This represents the probability of i values ​​occurring; The specific expression for edge strength is: （15）； （16）； （17）； （18）； （19）； In the formula, I represents the image grayscale matrix; "*" represents the convolution operation, which is the sum of the element-wise product of the neighborhood centered at (x, y) and the operator; G x and G y These represent the horizontal and vertical gradient operators in the Sobel algorithm, used to calculate the gradients in the horizontal and vertical directions, respectively; E x (x, y) and E y (x, y) represent using only G x and using only G y Perform binarization.

8. The multimodal fusion method for short-term load inflection point extraction and multi-segment prediction according to claim 3, characterized in that, The method for extracting important inflection point datasets using the important inflection point algorithm is as follows: A time series is a sequence of data points arranged chronologically, typically collected at equal intervals, with each data point representing a value at a specific time. A load curve time series is also included. S n Let be the load power at time n; definition S αm Let the m-th turning point satisfy the following conditions: （24）； （25）； （26）； （27）； In the formula, α m Representing point S m The change in angle at that location; θ m1 and θ m2 S represents m With adjacent S m-1 and S m+1 Angle value; Define the sequence of important turning points S iαk The k-th important turning point; three consecutive turning points S αj-1 S αj S αj+1 In the constructed sequence, S αj Can it become a major turning point and S αj To S αj-1 S αj+1 The distance D between the line segments is related to the distance S. The larger D is, the greater the distance S is. αj The farther away from the line segment, the more obvious the turning trend, and the greater the possibility of it becoming an important turning point; the vertical distance is used for calculation, and the formula is as follows. （28）； In the formula, v1, v2, and v2 represent three consecutive turning points S. αj-1 S αj S αj+1 Based on this formula, traverse the sequence of turning points and extract the important turning point sequence S. Iα .

Images