Power load prediction method, system and device and storage medium

Abstract

The invention discloses a power load prediction method, system and device, and a storage medium. The method comprises the steps of obtaining and cleaning power load historical data and environment feature data; the method comprises the following steps of: decomposing a plurality of intrinsic mode components by using generalized dynamic mode decomposition (GDMD), and dividing a training set and a test set; a GCN-GRELM fusion prediction model is constructed, and an improved triangular topology aggregation optimizer ITTAO is adopted to optimize parameters; calculating an adaptive weight based on the energy ratio of the subsequences and the sample entropy to obtain a preliminary prediction result; constructing an error data set by using the training set error sequence and the environment characteristic data, correcting a predicted value by combining an unscented Kalman filter (UKF) correction model, and outputting a final load prediction result; according to the method, the processing capability of nonlinear and non-stationary data is improved through GDMD decomposition and adaptive weight allocation, the generalization capability is enhanced in combination with a GCN-GRELM hybrid model and an ITTAO optimizer, and the calculation efficiency and prediction precision are improved by using GRELM efficient calculation and UKF dynamic correction.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system load forecasting, and particularly relates to a power load forecasting method, system, device and storage medium. Background Technique

[0002] Power load forecasting is a key link in power system planning and operation, and is widely used in power generation scheduling, power grid optimization, power market trading, new energy consumption and other fields. With the rapid development of smart grids and renewable energy, the operating environment of power systems has become increasingly complex, and the demand for refined load forecasting has been continuously increasing. At present, power enterprises, research institutions and technology suppliers are all actively exploring more efficient forecasting methods to improve forecasting accuracy and response speed to meet the high standards of modern power systems.

[0003] Existing power load forecasting methods are mainly divided into two categories: methods based on statistical models and methods based on machine learning. Statistical model methods mainly include time series analysis (such as ARIMA) and regression analysis, which rely on historical data for modeling and are suitable for load forecasting with obvious periodicity and trend. Machine learning methods use algorithms such as support vector machines (SVM) and artificial neural networks (ANN) to mine the non-linear laws of load changes through data training, improving the adaptability of forecasting.

[0004] Although existing methods can meet the requirements of power load forecasting to a certain extent, there are still limitations in the processing ability of non-linear and non-stationary data, limited model generalization ability, and low computational efficiency. Summary of the Invention

[0005] Object of the Invention: The object of the present invention is to provide a power load forecasting method that can improve the processing ability of non-linear and non-stationary data, the generalization ability of the model and the computational efficiency; on the other hand, to provide a power load forecasting system.

[0006] Technical Solution: The power load forecasting method described in the present invention includes the following steps:

[0007] (1) Obtain historical power load data and environmental characteristic data, and perform data cleaning to effectively remove noise and outliers, providing a high-quality and reliable data basis for subsequent analysis and improving the quality of the input data of the model;

[0008] (2) Use Generalized Dynamic Mode Decomposition (GDMD) to decompose the historical power load data and environmental characteristic data into several Intrinsic Mode Functions, and divide the training set and the test set according to a preset ratio, effectively capturing non-linear and non-stationary characteristics, and at the same time reasonably dividing the training set and the test set to ensure the scientific nature of model training and evaluation;

[0009] (3) Construct a fusion prediction GCN-GRELM model of a graph convolutional neural network and a generalized regularized extreme learning machine, and use the improved triangular topology aggregation optimizer ITTAO to optimize the parameters of the GCN-GRELM model, which can improve the convergence speed and generalization performance of the model;

[0010] (4) Input the historical power load data and environmental feature data into the optimized GCN-GRELM model to obtain several subsequence prediction results. Calculate the adaptive weights of the subsequences based on the energy ratio and sample entropy of each subsequence to obtain the preliminary prediction results, avoiding the bias of the traditional fixed-weight method and improving the accuracy of the preliminary prediction results;

[0011] (5) Subtract the preliminary prediction results from the true values of the training set to obtain the training set error sequence. Add the environmental feature data to the training set error sequence to construct the error data of the training set, which provides the key input for subsequent error correction and enhances the interpretability and correction ability of the model;

[0012] (6) Correct the preliminary prediction results through the error data of the training set and the unscented Kalman filter UKF error correction model, and output the final load prediction value, significantly improving the prediction accuracy and reliability.

[0013] Preferably, the environmental feature data described in step 1 includes temperature, humidity, solar irradiation angle, wind speed, precipitation, air pressure, and weather type. The data cleaning includes removing outliers and filling in missing values, effectively improving the integrity and reliability of the input data, providing more comprehensive and accurate input features for subsequent model training, and thus enhancing the adaptability and robustness of power load prediction to complex environmental factors.

[0014] Preferably, the decomposition formula of the GDMD described in step 2 is:

[0015]

[0016] where K is the number of dispersion modes; S i (f) is the i-th component; η(f) represents the noise signal, A i (f) is the time-varying amplitude of the component, τ i (λ) is the dispersion function, is the initial phase.

[0017] This design effectively separates the key dynamic features and noise signals in the data, thus significantly enhancing the analytical ability for the non-stationary and non-linear characteristics of power loads and laying a decomposition foundation for subsequent high-precision prediction.

[0018] Preferably, the node feature update formula of the GCN module described in step 3 is:

[0019]

[0020] where H l is the node feature representation of the l-th layer; is the graph degree matrix with self-loops added; is the graph adjacency matrix with self-loops added; W l is the trainable parameter matrix of the l-th layer; σ is the corresponding activation function, effectively extracting the spatial correlation features of the power load data.

[0021] The loss function of the GREML module is:

[0022]

[0023] where H is the feature matrix output by the GCN, β is the output weight matrix, y is the target label matrix, and λ is the regularization coefficient, jointly enhancing the modeling ability of the hybrid model for the complex spatio-temporal correlation features of the power system, preventing overfitting while improving the prediction accuracy, and ensuring the generalization performance of the model.

[0024] Preferably, the optimization process of ITTAO described in step 3 includes:

[0025] Initialization stage: Randomly generate N / 3 triangular topological units in the feasible region. The mathematical expression of the first search individual of each unit is:

[0026]

[0027] where represents the first search individual in the j-th triangular topological unit, and j is a positive integer between 1 and N / 3, r0 represents a random number between [0, 1], and are the lower and upper bounds of the variables;

[0028] Local aggregation stage: Taking the first vertex as the origin of the spherical coordinate system, generate a direction vector with a length of l*f, convert it to the ordinary coordinate system through trigonometric functions, and determine the second vertex, whose coordinates are:

[0029]

[0030] Rotate the direction vector counterclockwise by π / 3, and generate the third vertex through coordinate transformation, whose coordinates are:

[0031]

[0032] where l represents the size of the triangular topological unit, which is mathematically expressed as t represents the current iteration number, T represents the maximum number of iterations, l decreases as the number of iterations increases, and f(θ) and f(θ+π / 3) represent the direction vectors of the other two edges guided by the first point:

[0033]

[0034]

[0035] where θ=[θ1,...,θ D ] and θ o (o=1,...,D) is a random number between [0,π];

[0036] The fourth vertex is generated by linear weighted aggregation inside each triangle unit, and its coordinates are:

[0037]

[0038] Where r1, r2 and r3 are random numbers between [0, 1] and r1+r2+r3=1;

[0039] Global aggregation stage: Information interaction between different triangle units to generate new individuals:

[0040]

[0041] Where r4 is a random number between [0, 1], and represents the best individual of unit i and the randomly selected unit at iteration t;

[0042] Use greedy strategy to update the optimal solution of the population:

[0043]

[0044] in represents the suboptimal individual at the i-th iteration.

[0045] Through the multi-stage dynamic adjustment mechanism of the ITTAO optimizer, combined with the adaptively reduced search step size and iterative update of the greedy strategy, the global exploration and local development capabilities of the GCN-GRELM model in the parameter search process are effectively balanced, significantly improving the optimization efficiency and accuracy of the model parameters, while enhancing the convergence stability of the algorithm in complex non-convex spaces, thereby ensuring that the power load forecasting model can quickly obtain high-performance and generalization-strong parameter combinations.

[0046] Preferably, step 4 specifically includes:

[0047] Calculate the energy proportion E of each subsequence i and sample entropy S i, the specific formula is as follows:

[0048]

[0049] S i = SampleEntropy(IMF i )

[0050] where x i(t) is the value of the i-th subsequence at time t, T is the sequence length, SampleEntropy is the calculation formula of sample entropy, and IMF i is the intrinsic mode component;

[0051] Calculate the comprehensive weight of the subsequence according to the energy ratio and sample entropy. The mathematical expression is as follows:

[0052]

[0053] where α and β are adjustment parameters;

[0054] According to the comprehensive weight and the prediction result of the subsequence of the GRELM model obtain the final power load prediction result y i , and the formula is as follows:

[0055]

[0056] By dynamically calculating the energy ratio and sample entropy of each subsequence and adaptively allocating the comprehensive weight based on the adjustment parameters, the accurate quantitative evaluation of the key features of non-stationary power load data is realized, enabling the prediction model to effectively distinguish and strengthen the high-value mode components while suppressing noise interference, thereby significantly improving the accuracy and robustness of the prediction result, and finally outputting a more reliable power load prediction value through the weighted fusion mechanism.

[0057] Preferably, step 6 specifically includes:

[0058] (61) Generate a set of Sigma points according to the current state estimate and the covariance matrix Pt:

[0059]

[0060]

[0061]

[0062] where n is the state dimension, k is the scaling parameter used to adjust the distribution range of the Sigma points, represents the i-th column of the Cholesky decomposition of the matrix (n + k)P t ;

[0063] (62) Assign mean and covariance weights to each Sigma point:

[0064]

[0065]

[0066] Among them, is the mean weight, is the covariance weight, and α and β are adjustment parameters; (63) State prediction: Pass the Sigma points through the state transition function f(·):

[0067] x i|t = f(x i|t-1 ), i = 0, 1,..., 2n

[0068] Among them, x i|t and x i|t-1 are the sets of Sigma points at times t - 1 and t, respectively;

[0069] Calculate the predicted state mean and covariance:

[0070]

[0071]

[0072] Among them, Q t is the process noise covariance matrix;

[0073] (64) Observation prediction: Pass the Sigma points through the observation function h(·):

[0074] z i|t = h(x i|t-1 ), i = 0, 1,..., 2n

[0075] Among them, z i|t is the i-th observed point after mapping;

[0076] Calculate the predicted observation mean and covariance:

[0077]

[0078]

[0079] Among them, R t is the observation noise covariance matrix;

[0080] (65) State update: Calculate the cross-covariance matrix of the state and the observation:

[0081]

[0082] Calculate the Kalman gain:

[0083] K t = P xz P zz -1

[0084] Update the state estimate and covariance matrix:

[0085]

[0086]

[0087] where, z t is the actual observation value;

[0088] (66) Use the error data of the training set constructed in step 5 to train the UKF error correction model, use the trained UKF error correction model to predict the error of the test set to obtain the error prediction value of the test set, add the error prediction value of the test set to the preliminary prediction result obtained in step 4, obtain the final prediction result and output it.

[0089] Through the unscented transform (UT) process of the UKF error correction model, the nonlinear statistical characteristics of the prediction error are accurately captured by Sigma point sampling, combined with the dynamic weight allocation and state-observation dual-path correction mechanism, effectively reducing the model uncertainty; the UKF parameters optimized by the training set error data significantly improve the accuracy of error prediction, and finally the accuracy and stability of the power load prediction result reach the optimal through error compensation. This process fully integrates the historical error law and real-time observation information, enabling the system to have the ability of adaptive error suppression.

[0090] The power load prediction system described in the present invention includes:

[0091] A data acquisition and cleaning module, configured to acquire power load historical data and environmental characteristic data, and perform data cleaning;

[0092] A data decomposition and partitioning module, configured to decompose the power load historical data and environmental characteristic data into a plurality of intrinsic mode components using the generalized dynamic mode decomposition GDMD, and partition the training set and the test set according to a preset ratio;

[0093] A model construction and optimization module, configured to construct a graph convolutional neural network and a generalized regularized extreme learning machine fusion prediction GCN-GRELM model, and optimize the parameters of the GCN-GRELM model using an improved triangular topology aggregation optimizer ITTAO;

[0094] The preliminary structure prediction module is used to input the historical power load data and environmental feature data into the optimized GCN-GRELM model to obtain several subsequence prediction results, calculate the adaptive weights of the subsequences based on the energy ratio and sample entropy of each subsequence, and obtain the preliminary prediction results;

[0095] The error data construction module is used to subtract the preliminary prediction results from the true values of the training set to obtain the training set error sequence, and add the environmental feature data to the training set error sequence to construct the error data of the training set;

[0096] The error correction and output module is used to correct the preliminary prediction results through the error data of the training set and the unscented Kalman filter (UKF) error correction model, and output the final load prediction value.

[0097] A computer-readable storage medium, on which a computer program is stored, characterized in that when the computer program is executed by a processor, the described power load prediction method is implemented.

[0098] A computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that when the processor executes the computer program, the described power load prediction method is implemented.

[0099] Advantages: Compared with the prior art, the present invention has the following remarkable advantages: 1. It can improve the processing ability of non-linear and non-stationary data, the generalization ability of the model, and the calculation efficiency; 2. It can quantitatively evaluate the contribution degree of each subsequence to the power load prediction, dynamically allocate the weight coefficients of each component, avoid the problem of feature overfitting or underfitting caused by traditional fixed weights, and greatly improve the prediction accuracy; 3. The fast learning algorithm and regularization mechanism of GRELM can prevent overfitting while ensuring efficient training; 4. The UKF error dynamic correction further strengthens the model reliability, significantly improves the calculation efficiency and system robustness, and is applicable to the real-time prediction requirements of multiple scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0100] Figure 1 It is a flow chart of the present invention;

[0101] Figure 2 It is an overall architecture diagram of the present invention;

[0102] Figure 3 It is a schematic diagram of the GCN-GRELM hybrid model structure of the present invention;

[0103] Figure 4 It is a flow chart of the ITTAO algorithm of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0104] The technical solution of the present invention will be further described below in conjunction with the accompanying drawings.

[0105] As Figure 1 shown, the method includes the following steps:

[0106] Step 1: Pre-acquire the historical data of local power load. The historical data includes characteristic data such as temperature, humidity, solar irradiation angle, wind speed, precipitation, air pressure, weather, etc. and power load data; perform data cleaning on the data to prevent data errors caused by reasons such as unstable equipment and system failures.

[0107] Step 2: As Figure 2 shown, use GDMD to decompose the original data into k components, denoted as IMF1, IMF2, ……, IMFk; divide the training set and the test set for each subsequence according to the ratio of 7:3.

[0108] Use GDMD to decompose the original power load data into two parts: k components and a noise signal. The decomposition formula is as follows:

[0109]

[0110] Among them, K is the number of dispersion modes; S i (f) is the i-th component; η(f) represents the noise signal, A i (f) is the time-varying amplitude of the component, τ i (λ) is the dispersion function, is the initial phase.

[0111] Divide the decomposed subsequences into a training set and a test set according to the ratio of 7:3.

[0112] Step 3: Construct a fusion prediction model of a graph convolutional neuron and a generalized regularized extreme learning machine (GCN-GRELM);

[0113] (31) GCN is a deep learning network focusing on processing graph data. Its core idea is to generalize the convolution operation from the Euclidean space to the non-Euclidean space, and update the representation of each node by aggregating the information of the node and its neighbors, so that GCN can effectively learn the complex adjacency information relationship between different historical data; as Figure 3 shown, in the present invention, each acquired historical data is regarded as a node in the graph, and the connections between the historical data form the edges of the graph; the distance between different historical data is used as the feature of the edge, thus forming a graph data set for power load prediction; the calculation propagation process between the GCN network layers can be expressed as:

[0114]

[0115] Among them, Hl is the node feature representation of the l-th layer; is the graph degree matrix with self-loops added; is the graph adjacency matrix with self-loops added; W l is the trainable parameter matrix of the l-th layer; σ is the corresponding activation function.

[0116] (32) Different from the ELM model, the main feature of the GRELM model is that it adds structural risk on the basis of empirical risk. While ensuring the minimum training error, the structure of the model is simpler and the generalization performance is better; the node feature representation output by the GCN module is used as the input of the GRELM, and the GRELM learns the global feature representation by minimizing the regularized loss function:

[0117]

[0118] where H is the feature matrix output by the GCN, β is the output weight matrix, y is the target label matrix, and λ is the regularization coefficient.

[0119] Step 4: Based on the Triangular Topology Aggregation Optimizer TTAO, fuse multiple improvement strategies to obtain ITTAO, and use ITTOA to optimize the parameters of the GCN-GRELM model; the optimized model parameters include the regularization coefficient λ and the number of hidden layer neurons N;

[0120] (41) The original Triangular Topology Aggregation Optimizer TTAO has problems such as getting stuck in local optima and low convergence accuracy in some cases. In the present invention, the Triangular Topology Aggregation Optimizer TTAO is used to fuse multiple improvement strategies to obtain ITTAO to replace the original TTAO algorithm. As Figure 4 shown, in the initialization stage, N / 3 agents are randomly generated in the feasible region, and the mathematical expression generated by each individual is:

[0121]

[0122] where represents the first search individual in the i-th triangular topology unit, and i is a positive integer between 1 and N / 3, r0 represents a random number between [0, 1], and are the lower and upper bounds of the variables.

[0123] (42) Construct a new direction vector with a length of l*f and use the first vertex as the starting vertex in the spherical coordinate system for positioning, and convert it to the ordinary coordinate system through trigonometric functions to form the second vertex; then, the generated direction vector with a length of l*f is rotated counterclockwise by π / 3, and then the third vertex is obtained through coordinate transformation; the expressions of these vertices can be written as:

[0124]

[0125]

[0126] where l represents the size of the triangular topological unit, which is mathematically expressed as:

[0127]

[0128] where t represents the current iteration number, T represents the maximum number of iterations, and l decreases as the number of iterations increases. f(θ) and represent the direction vectors of the other two sides guided by the first point:

[0129]

[0130]

[0131] where θ = [θ1,..., θ D and θ j (j = 1,..., D) are random numbers between [0, π];

[0132] Each group of triangular topological units aggregates internally into a fourth vertex. This point is formed in a linearly weighted manner to use individual information, defined as:

[0133]

[0134] where r1, r2, and r3 are random numbers between [0, 1] and r1 + r2 + r3 = 1.

[0135] (43) In the global aggregation stage, the information of excellent individuals in different triangular units is collected, and new feasible solutions are created. Information interaction occurs between the best individual in each triangular topological unit and the best individual in an arbitrarily randomly selected unit set. There is a linear combination with different weights between each dimensional variable of the two positive individuals. The new individual is generated from the connection of the better two vertices, and the mathematical expression is:

[0136]

[0137] where r4 is a random number between [0, 1], and represent the best individual of unit i and the unit randomly selected at the t-th iteration; in addition, a greedy strategy is adopted to update the optimal agent, and the mathematical expression is:

[0138]

[0139] where represents the sub-optimal individual at the i-th iteration.

[0140] Step 5: Input the data into the optimized GCN-GRELM model to obtain the prediction results of K subsequences; define the adaptive weights of the subsequences according to the energy proportion and sample entropy of each subsequence, and obtain the preliminary prediction results based on the weights;

[0141] (51) Calculate the energy proportion E i and the sample entropy S i , and the specific formulas are as follows:

[0142]

[0143] S i = SampleEntropy(IMF i )

[0144] where x i(t) is the value of the i-th component at time t, T is the sequence length, and SampleEntropy is the calculation formula of the sample entropy;

[0145] Calculate the comprehensive weight of the component according to the energy proportion and sample entropy, and the mathematical expression is as follows:

[0146]

[0147] where α and β are adjustment parameters, α is 0.4, and β is 0.6.

[0148] (52) Obtain the final power load prediction result y according to the comprehensive weight and the prediction results of the GRELM model components i , and the formula is as follows:

[0149]

[0150] Step 6: Subtract the initial prediction result of the training set obtained in Step 5 from the true value of the training set to obtain the error sequence of the training set, add the original features to the error sequence, and use the features as the input to construct the error data of the original data.

[0151] Step 7: Construct an unscented Kalman filter (UKF) error correction model to obtain the error sequence of the test set, and add it to the preliminary prediction result obtained in Step 5 to obtain the final prediction result.

[0152] (71) UKF captures the mean and covariance of the state distribution by selecting a set of sample points called Sigma points and passes these points through the nonlinear system to obtain the updated state estimate. Compared with EKF, UKF can handle the state estimation problem of nonlinear systems more accurately. According to the current state estimate Generate a set of Sigma points with the covariance matrix Pt:

[0153]

[0154]

[0155]

[0156] where n is the state dimension and k is a scaling parameter used to adjust the distribution range of the Sigma points, denotes the i-th column of the Cholesky decomposition of the matrix (n + k)P t .

[0157] (72) Assign weights to each Sigma point:

[0158]

[0159]

[0160] where, is the mean weight, is the covariance weight, and α and β are adjustment parameters.

[0161] (73) State prediction: Pass the Sigma points through the state transition function f(·):

[0162] x i|t = f(x i|t-1 ), i = 0, 1,..., 2n

[0163] where x i|t and x i|t-1 are the sets of Sigma points at times t - 1 and t, respectively;

[0164] Calculate the predicted state mean and covariance:

[0165]

[0166]

[0167] where, Q t is the process noise covariance matrix.

[0168] (74) Observation prediction: Pass the Sigma points through the observation function h(·):

[0169] z i|t = h(x i|t-1 ), i = 0, 1,..., 2n

[0170] where z i|tThe i-th observed point after mapping;

[0171] Calculate the predicted observation mean and covariance:

[0172]

[0173]

[0174] where, R t is the observation noise covariance matrix.

[0175] (75) Update step:

[0176] Calculate the cross-covariance matrix of the state and the observation:

[0177]

[0178] Calculate the Kalman gain:

[0179] K t = P xz P zz -1

[0180] Update the state estimate and covariance matrix:

[0181]

[0182]

[0183] where, z t is the actual observed value.

[0184] (76) Fuse the training set error data obtained in step 6 with the original feature data to form power load error data; input the training set of the error data into the unscented Kalman filter (UKF) error correction model, and then perform test set error prediction through the trained UFK model to obtain a test set error sequence; add the obtained predicted error value to the test set prediction result obtained in step 5 to get the final prediction result.

[0185] The embodiment of the present invention also discloses a computer-readable storage medium.

[0186] Specifically, the computer-readable storage medium is used to store a computer program. When the computer program is executed by a processor, the methods in the above method embodiments are implemented. Those skilled in the art can understand that to implement all or part of the processes in the above method embodiments of the present application, it can be completed by instructing relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the above method embodiments. Among them, the storage medium can be a magnetic disk, an optical disc, a read-only memory (ROM), a random access memory (RAM), a flash memory, a hard disk drive (HDD), or a solid-state drive (SSD), etc.; the storage medium can also include a combination of the above types of memories.

[0187] An embodiment of the present invention also discloses a computer device.

[0188] Specifically, the computer device can be a desktop computer, a laptop computer, a handheld computer, a cloud server, or other computer devices. The computer device can include, but is not limited to, a processor and a memory. Among them, the processor and the memory can be connected through a bus or other means. Among them, the processor can be a central processing unit (CPU). The processor can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, graphics processing units (GPUs), embedded neural network processors (NPUs), or other dedicated deep learning coprocessors, discrete gate or transistor logic devices, discrete hardware components, etc. chips, or a combination of the above types of chips.

[0189] The memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the methods in the above embodiments of the present application. By running the non-transitory software programs, instructions, and modules stored in the memory, the processor can execute various functional applications and data processing of the processor, that is, implement the methods in the above method embodiments. The memory can include a program storage area and a data storage area. Among them, the program storage area can store control units and application programs required for at least one function; the data storage area can store data created by the processor and the like. In addition, the memory can include high-speed random access memory, and can also include non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state storage devices. In some embodiments, the memory may optionally include memories remotely provided relative to the processor, and these remote memories can be connected to the processor through a network. Examples of the above networks include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

Claims

1. A power load forecasting method, characterized in that, It includes the following steps: (1) Obtain historical power load data and environmental characteristic data, and perform data cleaning; (2) Use Generalized Dynamic Mode Decomposition (GDMD) to decompose the historical power load data and environmental characteristic data into several Intrinsic Mode Functions (IMFs), and divide the training set and the test set according to a preset ratio; (3) Construct a Graph Convolutional Neural Network and Generalized Regularized Extreme Learning Machine fusion prediction (GCN-GRELM) model, and use an Improved Triangular Topology Aggregation Optimizer (ITTAO) to optimize the parameters of the GCN-GRELM model; (4) Input the historical power load data and environmental characteristic data into the optimized GCN-GRELM model to obtain several subsequence prediction results, calculate the adaptive weights of the subsequences based on the energy ratio and sample entropy of each subsequence, and obtain the preliminary prediction results; (5) Subtract the preliminary prediction results from the true values of the training set to obtain the training set error sequence, add the environmental characteristic data to the training set error sequence, and construct the error data of the training set; (6) Correct the preliminary prediction results through the error data of the training set and the Unscented Kalman Filter (UKF) error correction model, and output the final load prediction value.

2. The power load forecasting method according to claim 1, wherein The environmental characteristic data described in step 1 includes temperature, humidity, solar irradiation angle, wind speed, precipitation, air pressure, and weather type. The data cleaning includes removing outliers and filling missing values.

3. The power load forecasting method according to claim 1, characterized in that, The decomposition formula of GDMD described in step 2 is: where K is the number of the dispersion modes; S i (f) is the i-th component; η(f) represents the noise signal, A i (f) is the time-varying amplitude of the component, τ i (λ) is the dispersion function, is the initial phase.

4. The power load forecasting method according to claim 1, characterized in that The node feature update formula of the GCN module described in step 3 is: where H l is the node feature representation of the l-th layer; is the graph degree matrix with self-loops added; is the graph adjacency matrix with self-loops added; W l is the trainable parameter matrix of the l-th layer; σ is the corresponding activation function. The loss function of the GREML module is: where H is the feature matrix output by the GCN, β is the output weight matrix, y is the target label matrix, and λ is the regularization coefficient.

5. The power load forecasting method according to claim 1, characterized in that The optimization process of ITTAO described in step 3 includes: Initialization stage: Randomly generate N / 3 triangular topology units in the feasible region. The mathematical expression of the first search individual of each unit is: Among them represents the first search individual in the j-th triangular topological unit, where j is a positive integer between 1 and N / 3, and r0 represents a random number between [0, 1]. and are the lower and upper bounds of the variables; Local aggregation stage: Taking the first vertex as the origin of the spherical coordinate system, generating a direction vector with a length of l*f, converting it to the ordinary coordinate system through trigonometric functions, and determining the second vertex, whose coordinates are: Rotate the direction vector counterclockwise by π / 3, and generate the third vertex through coordinate transformation. Its coordinates are: where \(l\) represents the size of the triangular topological unit, which is mathematically expressed as \(t\) represents the current iteration number, \(T\) represents the maximum number of iterations, \(l\) decreases as the number of iterations increases, and \(f(\theta)\) and \(f(\theta + \pi / 3)\) represent the direction vectors of the other two sides guided by the first point: where θ = [θ1,..., θ D and θ o (o = 1,..., D) are random numbers between [0, π]; Generate the fourth vertex through linear weighted aggregation inside each triangular unit. Its coordinates are: where r1, r2, and r3 are random numbers between [0, 1] and r1 + r2 + r3 = 1; Global aggregation stage: Perform information interaction between different triangular units to generate new individuals: where r4 is a random number between [0, 1], and denote the best individual of cell i and the cell randomly selected at the t-th iteration; Adopt a greedy strategy to update the population optimal solution: Among them represents the sub-optimal individual at the i-th iteration.

6. The power load forecasting method according to claim 1, wherein Step 4 specifically includes: Calculate the energy proportion E of each subsequence i and the sample entropy S i , and the specific formulas are as follows: S i = SampleEntropy(IMF i ) where x i(t) is the value of the i-th subsequence at time t, T is the sequence length, SampleEntropy is the calculation formula of sample entropy, and IMF i is the intrinsic mode component; Calculate the comprehensive weights of the subsequences according to the energy ratio and sample entropy. The mathematical expression is as follows: where α and β are adjustment parameters; According to the subsequence prediction results of the comprehensive weight and the GRELM model the final power load prediction result y is obtained i , and the formula is as follows:

7. The power load forecasting method according to claim 1, characterized in that Step 6 specifically includes: (61)According to the current state estimate and the covariance matrix Pt, generate a set of Sigma points: where n is the state dimension, k is the scaling parameter used to adjust the distribution range of the Sigma points, represents the i-th column of the Cholesky decomposition of the matrix (n + k)P t ; (62) Assign mean and covariance weights to each Sigma point: Among them, is the mean weight, is the covariance weight, and α and β are adjustment parameters; (63) State prediction: Pass the Sigma points through the state transition function f(·): x i|t = f(x i|t-1 ), i = 0, 1, ..., 2n where x i|t and x i|t-1 are the sets of Sigma points at times t-1 and t, respectively; Calculate the predicted state mean and covariance: Among them, Q t is the process noise covariance matrix; (64) Observation prediction: Pass the Sigma points through the observation function h(·): z i|t = h(x i|t-1 ), i = 0, 1, ..., 2n where z i|t is the i-th observation point after mapping; Calculate the predicted observation mean and covariance: where R t is the observation noise covariance matrix; (65) State update: Calculate the cross-covariance matrix between the state and the observation: Calculate the Kalman gain: K t = P xz P zz -1 Update the state estimate and covariance matrix: Among them, z t is the actual observed value; Use the error data of the training set constructed in step 5 to train the UKF error correction model, use the trained UKF error correction model to predict the error of the test set to obtain the error prediction value of the test set, add the error prediction value of the test set to the preliminary prediction result obtained in step 4, obtain the final prediction result and output it.

8. A power load forecasting system, characterized in that, Including: A data acquisition and cleaning module, which is used to acquire historical power load data and environmental feature data and perform data cleaning; A data decomposition and partitioning module, which is used to decompose the historical power load data and environmental feature data into several intrinsic mode components by using the generalized dynamic mode decomposition GDMD, and partition the training set and the test set according to a preset ratio; A model construction and optimization module, which is used to construct a graph convolutional neural network and a generalized regularized extreme learning machine fusion prediction GCN-GRELM model, and use the improved triangular topology aggregation optimizer ITTAO to optimize the parameters of the GCN-GRELM model; A preliminary structure prediction module, which is used to input the historical power load data and environmental feature data into the optimized GCN-GRELM model to obtain several subsequence prediction results, calculate the adaptive weights of the subsequences based on the energy ratio and sample entropy of each subsequence, and obtain the preliminary prediction result; An error data construction module, which is used to subtract the preliminary prediction result from the true value of the training set to obtain the training set error sequence, add the environmental feature data to the training set error sequence, and construct the error data of the training set; An error correction and output module, which is used to correct the preliminary prediction result through the error data of the training set and the unscented Kalman filter UKF error correction model, and output the final load prediction value.

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

10. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the power load prediction method described in any one of claims 1 to 7.

Images