Non-intrusive industrial load short-term prediction method based on diffusion model

Abstract

The invention discloses a non-intrusive industrial load short-term prediction method based on a diffusion model, and relates to the technical field of power industry energy management, and the method comprises the steps: converting power sequence data into a two-dimensional Grubrum angle field graph, and obtaining the converted image data; based on the image data, adopting a deep learning model to train and performing load decomposition on the new total port power data to obtain a load decomposition result; based on a load decomposition result, adopting a diffusion model to predict split-phase load power of a next time period; and based on the Grubrum angle field graph of the split-phase load power, carrying out Grubrum angle field inverse transformation on the two-dimensional image so as to obtain a power sequence prediction result of each load of the factory, and carrying out result analysis. According to the method, deep learning and a diffusion model are combined, efficient conversion and accurate decomposition of industrial load data are realized by innovatively combining a Grubrum angle field transformation method and the deep learning model, and then the diffusion model is utilized to predict power data of a next time period.

Description

Technical Field

[0001] The present invention relates to the technical field of energy management in the electric power industry, and in particular to a non-intrusive short-term industrial load forecasting method based on a diffusion model. Background Art

[0002] According to a report released by the National Energy Administration, by the end of June 2024, China's total installed power generation capacity was about 3.07 billion kilowatts, a year-on-year increase of 14.1%. Among them, the installed capacity of solar power generation was about 710 million kilowatts, a year-on-year increase of 51.6%; the installed capacity of wind power was about 470 million kilowatts, a year-on-year increase of 19.9%. In the first half of 2024, major power generation companies in China completed an investment of 344.1 billion yuan in power supply projects and 254 billion yuan in power grid projects, a year-on-year increase of 23.7%. The national power supply reached 3987.7 billion kilowatt-hours, and the cumulative average utilization hours of power generation equipment were 1666 hours, 71 hours less than the same period last year. In June 2024, the output of raw coal in industrial enterprises above designated size was 410 million tons, a year-on-year increase of 3.6%; the output of crude oil was 17.95 million tons, a year-on-year increase of 2.4%; and the output of natural gas was 20.2 billion cubic meters, a year-on-year increase of 9.6%. At the same time, the power generation of industrial enterprises above designated size was 768.5 billion kilowatt-hours, a year-on-year increase of 2.3%.

[0003] With the continuous growth of global energy demand and the proposed goal of large-scale access to renewable energy, the power system is facing unprecedented dual challenges and opportunities. As the largest and most diverse participant in the power system, the complexity and diversity of power loads have brought huge challenges to the operation and management of the power system.

[0004] Industrial load is a major component of the power system, and its stable operation is directly related to the development of the national economy and social stability. However, with the continuous advancement of industrialization, the growth momentum of industrial load continues unabated, which has brought severe challenges to the supply and demand balance and operational stability of the power system. At the same time, with the increase in the proportion of non-fossil energy and the large-scale access of renewable energy, the power system faces the challenge of how to effectively absorb intermittent renewable energy and how to ensure the safe and stable operation of the system. In this context, it is particularly important to build a data-based "source-grid-load-storage" interactive system. Among them, load monitoring technology, as a key link, can provide accurate and detailed data support for the operation and management of the power system. The development of load monitoring technology based on advanced measurement systems has attracted particular attention. It can obtain load information in real time and accurately without affecting the normal production and life of users, providing an important basis for the operation optimization and management of the power system. At the same time, accurate prediction of power load can better understand the power consumption characteristics and demand patterns of different loads, and realize the refined management and optimization of industrial loads, commercial loads and residential loads. This will help improve the operating efficiency and reliability of the power system, promote the efficient consumption of renewable energy, and lay a solid foundation for building a clean, intelligent and sustainable new power system.

[0005] In the field of non-intrusive load monitoring, most methods use advanced deep learning models for training, but there are generally limitations of single-perspective analysis. Although some studies use parallel network structures, the difference between the two branches in feature processing is not obvious, and there is a lack of in-depth discussion and analysis of the fusion of spatial and temporal features. In addition, the improvement of network structure and the combination of results are relatively simple, lacking diversified fusion strategies and theoretical basis, which limits the further improvement of model performance.

[0006] In the field of power data forecasting, many traditional forecasting models, such as the time series model ARIMA, require that the data is stable or stabilized after differentiation. They can only capture linear relationships and have limited ability to capture nonlinear patterns. Linear regression models are sensitive to outliers and may have overfitting problems, requiring model selection and parameter adjustment. LSTM training requires high hardware resources, large computational workload, and long training time. Summary of the invention

[0007] In view of the above-mentioned problems, the present invention is proposed.

[0008] Therefore, the technical problems solved by the present invention are:

[0009] To solve the above technical problems, the present invention provides the following technical solutions: a non-invasive short-term industrial load prediction method based on a diffusion model, comprising: transforming power series data into a two-dimensional Grammar angle field diagram to obtain transformed image data; based on the image data, using a deep learning model for training and performing load decomposition on the new total power data to obtain a load decomposition result; based on the load decomposition result, using a diffusion model to predict the phase load power in the next time period; based on the Grammar angle field diagram of the phase load power, performing an inverse Grammar angle field transform on the two-dimensional image, thereby obtaining the power series prediction results of each load in the factory, and performing result analysis.

[0010] As a preferred solution of the non-intrusive short-term industrial load forecasting method based on diffusion model described in the present invention, wherein: the power series data includes the total power data of the factory and the phase power data;

[0011] The step of transforming the power sequence data into a two-dimensional Grammar angle field image to obtain transformed image data refers to segmenting the power sequence data according to the plant load operation time characteristics, and performing a two-dimensional image transformation on the segmented power sequence data using a Grammar angle field transformation method, which specifically includes the following steps:

[0012] The power series data X of each group after segmentation i =[x1,x2,…,x N ] to normalize, X i The normalized value is recorded as

[0013] Will Converted to polar coordinates, the conversion formula is as follows:

[0014]

[0015] Among them, t i For x i timestamp; N is the number of time points contained in the power time series data; φ i is the polar angle; r i for The corresponding radius after transformation;

[0016] Calculate the power series data X i The inner product of gets the Gram matrix, and the formula is as follows:

[0017]

[0018] Among them, cos(φ i +φ j ) indicates X i Medium i With x jThe cosine of the sum of polar angles after polar coordinate conversion is calculated as follows:

[0019]

[0020] The corresponding pixels of the two-dimensional image are assigned values ​​according to the values ​​of the Gram matrix, thereby converting the power series data into a Gram angle field map.

[0021] As a preferred solution of the non-intrusive short-term industrial load forecasting method based on diffusion model described in the present invention, wherein: the forward process and backward process of the diffusion model;

[0022] The sampling process of the forward process satisfies the following formula:

[0023]

[0024] Among them, q(x t |x t-1 ) is x t The data at the moment is from x t-1 The sampling is obtained from The variance is β t , β t is a hyperparameter related to time t;

[0025] The sampling process of the reverse process satisfies the following formula:

[0026]

[0027] Among them, α t =1-β t , ∈ θ (x t ,t) is the noise data predicted by the neural network.

[0028] As a preferred solution of the non-intrusive short-term industrial load forecasting method based on the diffusion model described in the present invention, wherein: the Gram angle field map of the next time period output by the diffusion model is subjected to an inverse Gram transformation, and the data x of the pixels on the diagonal line of the Gram angle field map are first extracted. ii , and then calculate according to the following formula:

[0029]

[0030] The calculated data are spliced ​​together to obtain the power data of each load in the factory, realizing the prediction of industrial power data.

[0031] As a preferred solution of the non-intrusive short-term industrial load forecasting method based on diffusion model described in the present invention, wherein: the inverse process includes two parts: training and testing;

[0032] The training part includes the following steps:

[0033] Given raw data as input;

[0034] Randomly select time t from time step t = [1, 2, …, T];

[0035] Sample a noise ∈~N(0,I) from a standard Gaussian distribution;

[0036] Minimize the loss function by gradient descent in, is the noise predicted by the neural network;

[0037] Repeat the above process until the model training reaches convergence.

[0038] As a preferred solution of the non-intrusive short-term industrial load forecasting method based on diffusion model described in the present invention, the testing part includes the following steps:

[0039] Sample a noise x from a standard Gaussian distribution T ~N(0,I);

[0040] Starting from time step T, iterate backward to time step 1, and calculate the noise map corresponding to each time step t-1 according to the Gaussian distribution. The calculation formula is as follows:

[0041]

[0042] Among them, z ~ N (0, I);

[0043] When the iteration step is 1, z=0, and x0 is calculated, thereby completing the image prediction.

[0044] As a preferred solution of the non-intrusive short-term industrial load forecasting method based on diffusion model described in the present invention, wherein: the deep learning model includes a convolutional neural network CNN and a recurrent neural network RNN;

[0045] The convolutional neural network CNN is used to extract the pixel values ​​in the Gram angle field map to characterize the power amplitude of the industrial load as the spatial characteristic of the load.

[0046] The recurrent neural network RNN ​​is used to extract the time relationship between different Gram angle field diagrams to characterize the operating state of the industrial load as the time characteristic of the load;

[0047] The attention mechanism is introduced to assign different weights to different features of the load, distinguish different loads, and improve the training efficiency of the model.

[0048] To further solve the above technical problems, the present invention provides the following technical solutions: an industrial power prediction system based on a diffusion model and non-intrusive load monitoring, comprising: a data conversion module, used to convert power sequence data into a two-dimensional Grammar angle field diagram to obtain transformed image data; a load decomposition training module, used to train based on the image data using a deep learning model and perform load decomposition on the new total port power data to obtain a load decomposition result; a load prediction module, used to predict the phase load power of the next time period using a diffusion model based on the load decomposition result; a result inverse transformation and analysis module, used to perform a Grammar angle field inverse transformation on the two-dimensional image based on the Grammar angle field diagram of the phase load power, thereby obtaining the power series prediction results of each load in the factory, and performing result analysis.

[0049] A computer device includes a memory and a processor, wherein the memory stores a computer program, and is characterized in that when the processor executes the computer program, the steps of the non-intrusive industrial load short-term forecasting method based on the diffusion model as described above are implemented.

[0050] A computer-readable storage medium stores a computer program thereon, wherein when the computer program is executed by a processor, the steps of the non-intrusive short-term industrial load forecasting method based on a diffusion model as described above are implemented.

[0051] Beneficial effects of the present invention: The present invention combines deep learning and diffusion models, and through the innovative combination of the Gram angular field transformation method and the deep learning model, realizes the efficient conversion and precise decomposition of industrial load data, and then uses the diffusion model to predict the power data for the next time period. This method can not only significantly improve the accuracy of load forecasting, help factories and enterprises optimize the allocation of power generation resources and adjust power grid operation strategies, but also effectively implement demand-side management and reduce economic costs and environmental impacts. In addition, by accurately predicting electricity demand, enterprises can better integrate renewable energy, promote the green transformation of the energy structure, enhance the economy and sustainability of the entire energy system, and provide strong technical support for building a clean, efficient and intelligent energy environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0053] Figure 1 An overall flow chart of a non-intrusive short-term industrial load forecasting method based on a diffusion model provided by an embodiment of the present invention.

[0054] Figure 2 This is a neural network structure diagram of a non-intrusive load monitoring algorithm of the non-intrusive industrial load short-term forecasting method based on a diffusion model of the present invention.

[0055] Figure 3 It is a bidirectional gated cycle unit structure diagram of the non-intrusive short-term industrial load forecasting method based on the diffusion model of the present invention.

[0056] Figure 4 It is a diffusion model neural network structure diagram of the non-intrusive industrial load short-term forecasting method based on the diffusion model of the present invention.

[0057] Figure 5 This is a load decomposition result diagram of the non-invasive industrial load short-term forecasting method based on the diffusion model of the present invention.

[0058] Figure 6 This is a load forecasting result diagram of the non-intrusive industrial load short-term forecasting method based on the diffusion model of the present invention. DETAILED DESCRIPTION

[0059] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0060] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0061] Example 1, reference Figure 1 to Figure 4 , which is an embodiment of the present invention, provides a non-intrusive short-term industrial load forecasting method based on a diffusion model, comprising:

[0062] S1: transform the power sequence data into a two-dimensional Gram angular field image to obtain transformed image data.

[0063] It should be noted that this embodiment uses Shanxi Yangquan industrial data for case analysis. The sampling frequency of the data set is 1Hz, with total data for a period of 1 month and four types of industrial load data: cone crusher, press, crusher, and ball mill. The characteristics of the active power of the cone crusher are a peak of 30kW at startup, a stable operating power of about 3kW, and a fluctuation of about 0.5kW; the characteristics of the active power of the press are rapid start and stop, and the power switches quickly between a peak of about 7kW and 0W; the characteristics of the active power of the crusher are a peak of 30kW at startup, and the stable operating power slowly decreases from 2.5kW to 1.5kW, with small climbs and peaks appearing periodically; the characteristics of the active power of the ball mill are a peak of 75kW at startup, and the working power is about 27kW, with certain fluctuations. The data of the first 25 days of the data set are selected as the training set for load decomposition and load forecasting, and the data from the 26th to 29th days are used as the test set for load decomposition. The results of the load decomposition are used as the test set for load forecasting to predict the data of the last day and compare and analyze with the original data.

[0064] Further, transforming the power sequence data into a two-dimensional Grammar angle field image to obtain the transformed image data means segmenting the power sequence data according to the plant load operation time characteristics, and performing a two-dimensional image transformation on the segmented power sequence data using the Grammar angle field transformation method, which specifically includes the following steps:

[0065] The total port data and phase data are divided into hourly intervals, and each power sequence X i =[x1,x2,…,x N ] contains 3600 sampling points, N = 3600. The sequence is normalized according to the following formula:

[0066]

[0067] Among them, x min is the minimum value of the total port data, x max It is the maximum value of total port data.

[0068] The power series data X of each group after segmentation i =[x1,x2,…,x N ] to normalize, X i The normalized value is recorded as

[0069] Then calculate according to the formula The angles and radii of the sequence are converted to polar coordinate form:

[0070] Will Converted to polar coordinates, the conversion formula is as follows:

[0071]

[0072] Among them, t i Represents x i timestamp; N represents the number of time points contained in the power time series data, i.e. 3600. i The relationship between the original data values ​​is retained, r i The temporal relationship of the original data is preserved.

[0073] Calculate the power series data X i The inner product of gets the Gram matrix, and the formula is as follows:

[0074]

[0075] Among them, cos(φ i +φ j ) indicates X i Medium i With x j The cosine of the sum of polar angles after polar coordinate conversion is calculated as follows:

[0076]

[0077] The corresponding pixels of the two-dimensional image are assigned values ​​according to the values ​​of the Gram matrix, thereby converting the power series data into a Gram angle field map.

[0078] S2: Based on the image data, a deep learning model is used for training and load decomposition is performed on the new total port power data to obtain a load decomposition result.

[0079] First, build a deep learning neural network model, such as Figure 2 As shown in the figure, Conv1D represents a one-dimensional convolutional layer, which is used to calculate the dot product between the convolution kernel and the local area of ​​the input data. BiGRU represents a bidirectional gated recurrent unit, which is used to capture past and future information in the data. The convolutional neural network spatial feature extraction module consists of three one-dimensional convolutional layers, which are used to extract the size of the pixel value in the Gram angular field map, which can reflect φ i The relationship between and represents the amplitude of load power. The recurrent neural network temporal feature extraction module consists of three bidirectional gated recurrent units, which can reflect r i The relationship between represents the running time of the load. Further, the structure of BiGRU is as follows Figure 3 As shown, the formula of GRU is:

[0080] z t =σ(W Z ·[h t-1 ,x t ]);

[0081] rt =σ(W r ·[h t-1 ,x t ]);

[0082] h t =(1-z t )·h t-1 +z t ·tanh(W·[r t ·h t-1 ,x t ]);

[0083] Among them, z t Used to save historical data, t Used to update the input. W represents the weight matrix. The attention mechanism module consists of a weight matrix and weight features. It assigns different weights to different features, learns the importance of different parts of the input data, and generates richer and more discriminative feature representations.

[0084] Preferably, the BiGRU model processes sequence data through two independent GRU units in the forward and reverse directions, which can simultaneously capture the forward and backward information in the sequence, adaptively learn the long-term dependencies in the time series data and the mutual influence between multiple variables, and thus more comprehensively understand the patterns and trends in the sequence.

[0085] Then, the constructed model is used to train 25 days of industrial load data to obtain the Gram angle field diagram after load decomposition.

[0086] S3: Based on the load decomposition result, a diffusion model is used to predict the phase load power in the next time period.

[0087] First, a diffusion model is constructed. The diffusion model contains two Markov processes, namely the forward process and the reverse process.

[0088] The forward process means gradually adding random noise to the Gram angular field map to transform it into a pure noise image that obeys the Gaussian distribution. t From x t-1 The formula is:

[0089]

[0090] Among them, β t is a series of hyperparameters used to control the intensity of noise addition. According to the formula α t =1-β t as well as The above formula can be deduced to obtain:

[0091]

[0092] It can be concluded that the noise data x t From the mean The variance is is sampled from a Gaussian distribution.

[0093] The reverse process means denoising the image added with noise in the forward process and gradually restoring the predicted image. The sampling formula of the reverse process is as follows:

[0094]

[0095] Among them, the mean of the samples is the noise parameter ∈ output by the neural network θ (x t ,t) and cannot be directly derived.

[0096] Preferably, the goal of the neural network of the present invention is to predict the noise distribution at each step, rather than directly predicting the distribution of data. This method can reduce the difficulty of prediction because the distribution of noise is usually simpler than the distribution of data.

[0097] The reverse process is divided into two parts: training and testing. The training part is divided into the following steps:

[0098] (1) Given raw data as input;

[0099] (2) Randomly select time t from the time step t = [1, 2, …, T];

[0100] (3) Sample a noise ∈~N(0,I) from a standard Gaussian distribution;

[0101] (4) Minimize the loss function by gradient descent in, is the noise predicted by the neural network;

[0102] Preferably, (1) the form is simple, easy to optimize, and directly related to the prediction target of the model, that is, predicting the noise at each step, which simplifies the model training process. (2) It makes the training process more stable and is not prone to gradient explosion or vanishing problems.

[0103] (5) Repeat the above process until the model training reaches convergence.

[0104] The testing part is divided into the following steps:

[0105] (1) Sample a noise x from a standard Gaussian distribution T ~N(0,I);

[0106] (2) Starting from time step T, iterate backward to time step 1, and calculate the noise map corresponding to each time step t-1 according to the Gaussian distribution. The calculation formula is as follows:

[0107]

[0108] Among them, z~N(0,I).

[0109] Preferably, sampling from a distribution consisting of noise predicted by a neural network allows for more precise control of the generation process and faster sampling than sampling directly from a complex distribution.

[0110] (3) When the iteration step is 1, z=0, x0 is calculated, thus completing the image prediction.

[0111] In the reverse process, the neural network structure diagram is as follows Figure 4 As shown, the neural network has two inputs, one is the image data with noise added in the forward process, and the other is the time embedding corresponding to the image. ReLu, Switch, Tanh and Sigmoid are all activation functions, and the expressions are as follows:

[0112] ReLu: f(x) = max(0, x);

[0113]

[0114] At the same time, a multidimensional residual layer is introduced into the neural network to improve the problem that the gradient of each layer decreases rapidly after multiple multiplications, thereby improving the model's ability to fit the data.

[0115] S4: Based on the Gram angle field diagram of the phase-by-phase load power, perform an inverse Gram angle field transform on the two-dimensional image to obtain a power sequence prediction result of each load in the factory, and perform result analysis.

[0116] After obtaining the prediction results of the diffusion model output, it is necessary to convert the Gram angular field map into power series data. First, extract the pixel values ​​of the diagonal elements of the Gram angular field map, and then calculate according to the formula:

[0117]

[0118] The obtained x i The data can be calculated using the following formula to obtain the real power series data:

[0119]

[0120] Among them, x max Represents the maximum value of the original total port power sequence, x min Indicates the minimum value of the original total port power sequence, which is 0 in this case.

[0121] In order to evaluate the performance of the non-intrusive industrial load short-term forecasting method based on the diffusion model of the present invention, the accuracy, precision, recall and F1-score are selected as the performance indicators of the non-intrusive load algorithm proposed in step 2, and the mean square error (MSE) and root mean square error (RMSE) are selected as the performance indicators of the load forecasting algorithm proposed in step 3. The formula is as follows:

[0122]

[0123] Among them, TP is the number of true positive examples, TN is the number of true negative examples, FP is the number of false positive examples, and FN is the number of false negative examples.

[0124]

[0125] Among them, Y i is the actual value, is the predicted value and n is the number of samples.

[0126] In summary, the present invention combines deep learning and diffusion models, and adopts a fusion deep learning model to comprehensively capture the spatial and temporal characteristics of the load usage process. By adaptively assigning weights to different features, the model's response to useful information is enhanced, while reducing the interference of noise features. At the same time, the diffusion model is used in combination with a structured state space model to improve the architecture of the diffusion model, which is particularly suitable for capturing long-term dependencies in time series data, further improving the performance of load forecasting.

[0127] Embodiment 2 is an embodiment of the present invention, which provides an industrial power forecasting system based on diffusion model and non-intrusive load monitoring, including: a data transformation module, a load decomposition training module, a load forecasting module, and a result inverse transformation and analysis module.

[0128] The data conversion module is used to convert the power sequence data into a two-dimensional Gram angle field diagram to obtain the transformed image data.

[0129] The load decomposition training module is used to train the deep learning model based on image data and perform load decomposition on the new total port power data to obtain the load decomposition results.

[0130] The load forecasting module is used to predict the phase load power in the next time period based on the load decomposition results using a diffusion model.

[0131] The result inverse transformation and analysis module is used to perform inverse Gram angle field transformation on the two-dimensional image based on the Gram angle field diagram of the phase load power, so as to obtain the power sequence prediction results of each load in the factory and perform result analysis.

[0132] Embodiment 3 is an embodiment of the present invention, which is different from the previous embodiment in that: if the function is implemented in the form of a software function unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0133] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0134] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0135] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0136] Example 4, reference Figure 5 and Figure 6 , which is an embodiment of the present invention, provides a non-intrusive short-term industrial load forecasting method based on a diffusion model. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through performance evaluation analysis of the economic benefit calculation algorithm and simulation experiments.

[0137] Figure 5 Table 1 shows the results of some non-intrusive load decomposition algorithms. The evaluation indicators of the algorithm are shown in Table 1. Overall, the proposed non-intrusive load decomposition algorithm shows high accuracy, recall precision and F1 score in all test categories, which shows that the algorithm has high prediction accuracy and recognition ability for different devices. However, the F1 score of the press is low, which may be related to the excessive oscillation of the press waveform.

[0138] Figure 6 Table 2 is the result of load forecasting based on the diffusion model, and the evaluation index of the algorithm is shown in Table 2. Compared with the scale of each load power amplitude that may reach thousands or even tens of thousands, the load forecasting algorithm based on the diffusion model shows superior performance, and its root mean square error and relative average error results are both below 1500. This result shows that although the load forecast faces large volatility, the model can still provide relatively accurate forecasts, and the error is controlled at a relatively low level, which is of great significance for ensuring the stability and efficiency of the production process.

[0139] Table 1 Load decomposition evaluation indicators

[0140]

[0141] Table 2 Load forecasting evaluation indicators

[0142]

[0143] From the analysis results, it can be seen that the proposed load decomposition algorithm and load forecasting algorithm have achieved ideal results. By accurately identifying and predicting changes in electricity demand, they provide important support for industrial optimization of energy distribution, reduction of operating costs, enhancement of grid stability, promotion of renewable energy integration, support for demand-side management, risk assessment and improvement of user satisfaction, thereby promoting the development of smart grids and the efficient operation of the power market.

[0144] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A non-intrusive short-term industrial load forecasting method based on a diffusion model, characterized in that: include: Transform the power sequence data into a two-dimensional Gram angular field diagram to obtain transformed image data; Based on the image data, a deep learning model is used for training and load decomposition is performed on the new total port power data to obtain a load decomposition result; Based on the load decomposition result, a diffusion model is used to predict the phase load power in the next time period; Based on the Gram angle field diagram of the phase-by-phase load power, an inverse Gram angle field transform is performed on the two-dimensional image to obtain the power sequence prediction results of each load in the factory, and the results are analyzed.

2. The non-intrusive short-term industrial load forecasting method based on diffusion model according to claim 1, characterized in that: The power sequence data includes factory total power data and phase power data; The step of transforming the power sequence data into a two-dimensional Grammar angle field image to obtain transformed image data refers to segmenting the power sequence data according to the plant load operation time characteristics, and performing a two-dimensional image transformation on the segmented power sequence data using a Grammar angle field transformation method, which specifically includes the following steps: The power series data X of each group after segmentation i =[x1, x2, ..., x N ] to normalize, X i The normalized value is recorded as Will Converted to polar coordinates, the conversion formula is as follows: Among them, t i For x i timestamp; N is the number of time points contained in the power time series data; φ i is the polar angle; r i for The corresponding radius after transformation; Calculate the power series data X i The inner product of gets the Gram matrix, , the formula is as follows: Among them, cos(φ i +φ j ) indicates X i Medium i With x j The cosine of the sum of polar angles after polar coordinate conversion is calculated as follows: The corresponding pixels of the two-dimensional image are assigned values ​​according to the values ​​of the Gram matrix, thereby converting the power series data into a Gram angle field map.

3. The non-intrusive short-term industrial load forecasting method based on diffusion model according to claim 2, characterized in that: The forward process and backward process of the diffusion model; The sampling process of the forward process satisfies the following formula: Among them, q(x t |x t-1 ) is x t The data at the moment is from x t-1 The sampling is obtained from The variance is β t , β t is a hyperparameter related to time t; The sampling process of the reverse process satisfies the following formula: Among them, α t =1-β t , ∈ θ (x t , t) is the noise data predicted by the neural network.

4. The non-intrusive short-term industrial load forecasting method based on diffusion model according to claim 3, characterized in that: Perform an inverse Gram transformation on the Gram angular field map of the next time period output by the diffusion model. First, extract the data x of the pixels on the diagonal line of the Gram angular field map. ii , and then calculate according to the following formula: The calculated data are spliced ​​together to obtain the power data of each load in the factory, realizing the prediction of industrial power data.

5. The non-intrusive short-term industrial load forecasting method based on diffusion model according to claim 4, characterized in that: The reverse process includes two parts: training and testing; The training part includes the following steps: Given raw data as input; Randomly select time t from time steps t = [1, 2, ..., T]; Sample a noise ∈~N(0,I) from a standard Gaussian distribution; Minimize the loss function by gradient descent in, is the noise predicted by the neural network; Repeat the above process until the model training reaches convergence.

6. The non-intrusive short-term industrial load forecasting method based on diffusion model according to claim 5, characterized in that: The testing part includes the following steps, Sample a noise x from a standard Gaussian distribution T ~N(0,I); Starting from time step T, iterate backward to time step 1, and calculate the noise map corresponding to each time step t-1 according to the Gaussian distribution. The calculation formula is as follows: Where, z ~ N (0, I); When the iteration step is 1, z=0, and x0 is calculated, thereby completing the image prediction.

7. The non-intrusive short-term industrial load forecasting method based on diffusion model according to claim 6, characterized in that: The deep learning model includes a convolutional neural network (CNN) and a recurrent neural network (RNN); The convolutional neural network CNN is used to extract the pixel values ​​in the Gram angle field map to characterize the power amplitude of the industrial load as the spatial characteristic of the load. The recurrent neural network RNN ​​is used to extract the time relationship between different Gram angle field diagrams to characterize the operating state of the industrial load as the time characteristic of the load; The attention mechanism is introduced to assign different weights to different features of the load, distinguish different loads, and improve the training efficiency of the model.

8. A system using the non-intrusive short-term industrial load forecasting method based on a diffusion model as claimed in any one of claims 1 to 7, characterized in that: include: A data conversion module, used for converting the power sequence data into a two-dimensional Gram angle field diagram to obtain transformed image data; A load decomposition training module, used to train the image data using a deep learning model and perform load decomposition on the new total port power data to obtain a load decomposition result; A load prediction module, used to predict the phase load power in the next time period by using a diffusion model based on the load decomposition result; The result inverse transformation and analysis module is used to perform a Gram angle field inverse transformation on the two-dimensional image based on the Gram angle field diagram of the phase load power, so as to obtain the power sequence prediction results of each load in the factory and perform result analysis.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the non-intrusive short-term industrial load forecasting method based on a diffusion model according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the non-intrusive short-term industrial load forecasting method based on a diffusion model according to any one of claims 1 to 7 are implemented.

Images