Transformer area short-term load prediction method based on integrated diffusion filtering and improved graph convolution
By integrating diffusion filtering and improved graph convolution, the problems of noise filtering and nonlinear fluctuations in low-voltage distribution area load forecasting are solved, achieving high-precision load forecasting and supporting fine-grained scheduling of low-voltage distribution areas.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-08
- Publication Date
- 2026-03-17
AI Technical Summary
Existing load forecasting technologies lack a refined preprocessing mechanism in low-voltage distribution areas. Noise and outliers are not effectively filtered out, and it is difficult to capture the complex non-stationarity and nonlinear fluctuations of load sequences, resulting in insufficient forecast accuracy and inability to meet the needs of refined scheduling and control.
Noise is removed using integrated diffusion filtering technology. Graph structure data is constructed, and load forecasting is performed using improved graph convolutional layers, gated recurrent unit layers, and an improved Kolmogorov-Arnold network. The optimal model parameters are selected using the SMART method.
It significantly improves the quality of model input data, enhances the ability to fit complex nonlinear load fluctuations, improves prediction accuracy and stability, and can accurately capture peak load characteristics to meet the fine-grained scheduling needs of low-voltage distribution areas.
Smart Images
Figure CN121688844A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of short-term load forecasting technology, specifically to a method for short-term load forecasting of transformer areas based on integrated diffusion filtering and improved graph convolution. Background Technology
[0002] As a crucial link connecting the distribution network and users, the accuracy of load forecasting for low-voltage distribution areas directly impacts the network's dispatching decisions and economic operation. Achieving accurate load forecasting for low-voltage distribution areas has become an important technical foundation for improving the intelligence level of the distribution network and ensuring high-quality power supply for users.
[0003] Current research on load forecasting technology mainly focuses on high-voltage and medium-voltage levels of the power grid or large aggregated loads, while research on forecasting methods specifically for low-voltage distribution areas is relatively scarce. In practical applications, existing technologies mainly suffer from the following shortcomings: First, they lack a refined preprocessing mechanism for the characteristics of low-voltage distribution area data, and the noise and outliers contained in the raw load data are not effectively filtered out, directly affecting the quality of model input. Second, due to the low aggregation degree and strong randomness of user electricity consumption behavior in low-voltage distribution areas, their load sequences exhibit more severe non-stationarity and nonlinear fluctuations, and existing forecasting models cannot fully capture their complex internal variation patterns, resulting in insufficient forecasting accuracy and poor adaptability, failing to meet the engineering requirements for refined scheduling and control of low-voltage distribution areas. Summary of the Invention
[0004] This invention provides a method for short-term load forecasting of low-voltage distribution areas based on integrated diffusion filtering and improved graph convolution, aiming to solve the problems of strong load fluctuation and insufficient forecasting accuracy in low-voltage distribution areas.
[0005] To achieve the above objectives, the present invention provides the following technical solution: This invention relates to a short-term load forecasting method for transformer substations based on integrated diffusion filtering and improved graph convolution, comprising: S100: Acquire electrical and environmental characteristic data of the target low-voltage distribution area to form the original characteristic dataset; S200: An integrated diffusion filtering technique is proposed to remove noise and standardize the original feature dataset to obtain a standardized dataset. Each time step is used as a graph node, and a one-way edge adjacency matrix is constructed to describe the connection relationship between nodes, thus modeling the power load time series data as graph structure data. S300: Construct a hybrid deep learning prediction model, which includes graph convolutional layers, gated recurrent unit layers, and the output layer of an improved Kolmogorov-Arnold network. The computation process is reconstructed to activate the input with different basis functions and then perform a linear combination. S400: Input the standardized dataset and the one-way edge adjacency matrix into the hybrid deep learning prediction model for iterative training. Output the load prediction value and record multiple performance index values in each training round. S500: The SMART method is used to standardize and weight multiple performance index values, select the optimal model parameters based on the comprehensive evaluation index values, and use the optimal model parameters to perform short-term load forecasting of the transformer area based on the data to be predicted.
[0006] As a preferred embodiment of the present invention, the original feature dataset includes: Electrical characteristic data includes active power, reactive power, voltage value, and current value; Environmental characteristics data include temperature, humidity, wind speed, rainfall, and holiday markers.
[0007] As a preferred embodiment of the present invention, the step of the integrated diffusion filter removing noise from the original feature dataset includes: In the forward process, the original load sequence is copied multiple times, and Gaussian noise, uniform noise and Poisson noise are gradually added to the sequence to gradually transition it from clean data to a high noise state, thereby forming sequence representations with different noise levels and different noise types. In the reverse process, a trained model is used to gradually remove noise starting from the high-noise representation, and the signal structure is restored step by step in the opposite direction to the forward process, resulting in multiple smooth load sequences. By utilizing the concept of ensemble processing, the optimal combination of different noise diffusion models is achieved, ultimately yielding the denoised load sequence.
[0008] As a preferred embodiment of the present invention, the standardization process employs a maximum-minimum standardization method to process the denoised dataset.
[0009] As a preferred embodiment of the present invention, the one-way edge adjacency matrix only allows information to be transmitted unidirectionally from the preceding time step node to the subsequent time step node.
[0010] As a preferred embodiment of the present invention, the hybrid deep learning prediction model includes: The graph convolutional layer takes a standardized dataset and a one-way edge adjacency matrix as input, and performs feature aggregation based on the node connection relationships defined by the adjacency matrix; The gated loop unit layer includes an update gate and a reset gate. The update gate controls the retention ratio of the hidden state in the previous time step, and the reset gate controls the forgetting ratio of the hidden state in the previous time step. The output layer of the Kolmogorov-Arnold network uses B-spline functions to achieve parameterization through a linear combination of B-spline basis functions and trainable coefficients. Furthermore, its computational structure is reconstructed by changing L1 regularization based on input samples to L1 regularization based on weights, activating the input with different basis functions, and then performing a linear combination.
[0011] As a preferred embodiment of the present invention, the step of cyclic training includes: setting an initial training round, gradually increasing the number of training rounds for cyclic training, until the preset maximum number of training rounds is reached.
[0012] As a preferred embodiment of the present invention, the multiple performance index values include mean absolute error, mean absolute percentage error, root mean square error, and coefficient of determination.
[0013] As a preferred embodiment of the present invention, the standardization and weight fusion of multiple performance index values using the SMART method includes: The positive and negative performance indicators are standardized separately. Assign temporary weights to each performance metric to improve it from the worst to the best state; Normalize the temporary weights to obtain the weight coefficients of each performance index; The standardized performance index values recorded in each training round are weighted and summed with the corresponding weight coefficients to obtain the comprehensive evaluation index value for each training round. The model parameters corresponding to the training round with the largest comprehensive evaluation index value are selected as the optimal model parameters.
[0014] The beneficial effects of this invention are: 1. This invention addresses the severe noise interference in low-voltage distribution area data by proposing an integrated diffusion filtering technique. This technique introduces three different types of diffusion processes—Gaussian noise, uniform noise, and Poisson noise—and integrates the denoising results of multiple diffusion models using an ensemble strategy to achieve robust noise removal. This method can remove spike noise while preserving the periodicity and trend characteristics of the load sequence, significantly improving the quality of the model input data.
[0015] 2. This invention uses a one-way edge adjacency matrix to model time series data as a graph structure, re-examining time series data from a new perspective, and achieving information interaction and aggregation of features in the time dimension through graph convolutional layers. This operation enables the model to enhance the periodic and trend features in the original data with relatively low training cost.
[0016] 3. This invention uses the Kolmogorov-Arnold network as the output layer and reconstructs its computational structure, changing L1 regularization based on input samples to L1 regularization based on weights, activating the inputs with different basis functions, and then performing a linear combination. This reconstruction significantly reduces memory overhead, transforms computation into simple matrix multiplication, and naturally accommodates both forward and backward propagation. It significantly enhances the model's ability to fit complex nonlinear fluctuations in low-voltage distribution area loads. This network can more accurately capture peak load characteristics caused by disordered charging of electric vehicles, simultaneous activation of air conditioners, etc., improving prediction accuracy and stability. Attached Figure Description
[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the short-term load forecasting method for transformer areas based on integrated diffusion filtering and improved graph convolution of the present invention. Figure 2 This is a detailed flowchart illustrating the training and evaluation process of the hybrid deep learning model of this invention. Figure 3 This is a schematic diagram illustrating the effect of the proposed integrated diffusion filtering technique; Figure 4 This is a comparison chart of the prediction curves in this invention. Detailed Implementation
[0018] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.
[0019] Example 1: As Figure 1 As shown, the present invention provides a method for short-term load forecasting of transformer substations based on integrated diffusion filtering and improved graph convolution, comprising: S100: Acquire electrical and environmental characteristic data of the target low-voltage distribution area to form the original characteristic dataset; Furthermore, the original feature dataset includes: Electrical characteristic data includes active power, reactive power, voltage value, and current value; Environmental characteristics data include temperature, humidity, wind speed, rainfall, and holiday markers.
[0020] Specifically, during the data acquisition phase, the raw feature dataset X for short-term load forecasting is obtained from the target low-voltage distribution area.
[0021] Specifically, electrical characteristic data is acquired through distribution automation systems, distribution management systems, and electricity consumption information collection systems, including: active power flowing on the transformer busbars. (Unit: kWh) Reactive power (Unit: kWh), Voltage Value (Unit: V) and current value (Unit: A). Among them, active power... The target characteristic to be predicted is the power load value of the low-voltage distribution area.
[0022] Environmental characteristic data is acquired through intelligent distribution transformer terminals, integrated environmental monitoring sensors, and the meteorological data system of the target transformer area, including: temperature. (Unit: °C), Humidity (Unit: %) Wind speed (Unit: m / s) Rainfall (Unit: L / m²) and holiday signs (1 for yes, 0 for no).
[0023] The above 9 columns of feature data are sorted by column to form the original feature dataset. This ensures that the collection frequency of all feature columns remains consistent with the total number of samples, providing basic data support for subsequent data preprocessing and model training.
[0024] S200: An integrated diffusion filtering technique is proposed to remove noise and standardize the original feature dataset, resulting in a standardized dataset. Each time step is used as a graph node, and a one-way edge adjacency matrix is constructed to describe the connection relationship between nodes. Specifically, first, all missing values are discarded. Then, the mean of each feature column is calculated without considering missing values. Outliers greater than three times the mean are filtered out and replaced with their missing values. Next, linear interpolation is used to fill in all missing values, calculated using the following formula: ; in, For the previous moment, For the next moment, For the time corresponding to the missing value, Given the known characteristic values from the previous time step. Given the known eigenvalues at the next time step. This represents the missing interpolation result.
[0025] Furthermore, the step of using the integrated diffusion filtering technique to remove noise from the original feature dataset includes: In the forward process, the original load sequence is copied multiple times, and Gaussian noise, uniform noise and Poisson noise are gradually added to the sequence to gradually transition it from clean data to a high noise state, thereby forming sequence representations with different noise levels and different noise types. In the reverse process, the trained model is used to gradually remove noise starting from the high-noise representation. The signal structure is restored step by step in the opposite direction to the forward process, resulting in multiple smooth load sequences. These are then averaged and integrated to finally obtain the denoised load sequence.
[0026] Specifically, let the clean power load time series be... The observation sequence is contaminated by noise, denoted as Integrated diffusion filtering technology learns to recover the original sequence from noise through a three-stage process of "multiple noise addition - noise reduction - integration".
[0027] During the forward pass, the original load sequence data is first copied three times. Then, Gaussian noise, uniform noise, and Poisson noise are gradually added to the three copies of the data, respectively.
[0028] Let the total number of diffusion steps be Time step The noise intensity scheduling parameters are ,definition The forward diffusion process is as follows: ; in, Let be the identity matrix, representing a given hour, Follow the mean variance is The Gaussian distribution.
[0029] We can obtain the form of directly sampling from the original sequence to any diffusion step: ; in, It is standard Gaussian noise.
[0030] when When large enough, Approximately isotropic Gaussian noise.
[0031] The goal is to learn the reverse process. This makes the noise Stepwise sampling can reconstruct .in This represents the parameters of the neural network to be trained.
[0032] Typically, noise prediction parameterization is used, which involves training a network. Estimating the noise added during forward diffusion The training loss is: ; in, This indicates the expected operation.
[0033] Using the estimated noise, the mean of the reverse one-step sampling can be constructed: ; Therefore, the inverse distribution takes Gaussian form: ; in This is to derive the effective variance.
[0034] By from Start sampling sequentially: ; This will give you the Gaussian diffusion denoising sequence. The subscript 1 indicates the first type of noise (Gaussian).
[0035] Similarly, for uniform diffusion, in the forward process, the original load sequence In each step of the injected interval The uniform noise on the surface gradually transitions to a noise distribution, forming a noisy sequence.
[0036] The forward diffusion process is as follows: ; in For noise sampled from a uniform distribution, The noise amplitude is controlled to increase with each step.
[0037] During the reverse engineering process, the training network is utilized. Estimate the uniform noise component in each step, and recover the noisy sequence in reverse order to obtain: ; Gradually eliminate uniform noise disturbances of different scales and reconstruct a smooth load sequence based on uniform noise diffusion. .
[0038] Similarly, for Poisson diffusion, in the forward process, the original load sequence At each step, a Poisson distribution is randomly perturbed, gradually transitioning it towards a noise-dominated distribution to form a noisy sequence: ; in The Poisson noise intensity is controlled as the steps increase.
[0039] During the reverse engineering process, the training network is utilized. Estimate the Poisson noise component in each step and recover the noisy sequence in reverse order to obtain: ; Gradually eliminate the random bias introduced by Poisson disturbances to reconstruct a smooth load sequence. .
[0040] To combine the advantages of different noise diffusion models in terms of denoising effect and improve the robustness and diversity of denoising results, a weighted integration of their output sequences is performed.
[0041] First, a confidence function is defined based on the residual information of the input for each model. For example, it can be obtained from the prediction residuals or the estimation of network internal uncertainties: ; in, For a monotonically decreasing function, the smaller the residual, the larger the weight. Corresponding to Gaussian, uniform, and Poisson diffusion models respectively, and obtained through normalization: ; The final integration result (denoted as) ,in (Indicating ensemble integration) is represented as: ; Furthermore, the standardization process employs a maximum-minimum standardization method to process the denoised dataset.
[0042] Specifically, the denoised dataset is standardized. For each feature sequence, the max-min standardization method is used to eliminate the influence of differences in the units of measurement between different features. The calculation formula is: ; in, It is the normalized value. It is the raw data. and These represent the maximum and minimum values in the original feature data column, respectively.
[0043] Furthermore, the one-way edge adjacency matrix only allows information to be transmitted unidirectionally from the preceding time step node to the subsequent time step node.
[0044] Specifically, the standardized time-series data is modeled as graph-structured data. Graph convolutional neural networks require edges between nodes in the input data to represent the relationships between them. Therefore, before feeding the data into the model, an adjacency matrix A is needed to describe the topological relationships between nodes. Its calculation formula is as follows: ; ; in, The length of the time series. Adjacency matrix The elements in the adjacency matrix are arranged in a grid. An element with a value of 1 indicates an edge between two nodes, while a value of 0 indicates the opposite. Considering the practical significance of time nodes, information from later time nodes should not be passed to earlier time nodes. Therefore, all edges in the constructed graph data are unidirectional, meaning information is only allowed to flow unidirectionally from previous time step nodes to subsequent time step nodes.
[0045] Time series data is modeled as graph-structured data by using an adjacency matrix. Each time step in each sample corresponds to a node in the graph, and all feature data in each time step corresponds to the feature value of the current node, so as to exchange information in subsequent graph convolutional layers.
[0046] S300: Construct a hybrid deep learning prediction model, which includes graph convolutional layers, gated recurrent unit layers, and the output layer of an improved Kolmogorov-Arnold network. Furthermore, the hybrid deep learning prediction model includes: The graph convolutional layer takes a standardized dataset and a one-way edge adjacency matrix as input, and performs feature aggregation based on the node connection relationships defined by the adjacency matrix; The gated loop unit layer includes an update gate and a reset gate. The update gate controls the retention ratio of the hidden state in the previous time step, and the reset gate controls the forgetting ratio of the hidden state in the previous time step. The output layer of the Kolmogorov-Arnold network uses B-spline functions to parameterize the input through a linear combination of B-spline basis functions and trainable coefficients, and reconstructs its computational structure by activating the input with different basis functions and then performing a linear combination.
[0047] Specifically, the graph convolutional layer receives a standardized dataset and a one-way edge adjacency matrix as input. The modeled graph structure data consists of a set of nodes V and a set of edges E. The input feature matrix H is composed of the features of all nodes, and the adjacency matrix A represents the edges connecting the nodes, which physically represent the connections between nodes that allow information transfer. H and A are used as input to the graph convolutional layer, and the output Y is obtained through the processing function f of the graph convolutional layer. ; The convolution feedforward method for each layer of the graph convolutional layer is as follows: ; in, It is the first Input to the layer network, ; The number of nodes in the graph, each node uses 3D eigenvector representation; Let be the adjacency matrix of an undirected graph. for An identity matrix of order 1; Let be the degree matrix of an undirected graph. ; These are the parameters to be trained; For output dimensions; The corresponding activation function is . Graph convolutional layers aggregate features through node connections defined by the adjacency matrix, directly passing information about the current node and indirectly passing information about the preceding nodes to extract features from the data, making the trends between the data more obvious.
[0048] The gated loop unit layer includes two core gating mechanisms: the update gate and the reset gate.
[0049] The update gate controls the proportion of the hidden state retained from the previous time step, determining whether the current hidden state relies more on past information or new input information: ; in, It updates the output of the gate. It is an activation function. It updates the weight matrix of the gate. It is the hidden state of the previous time step. This is the input for the current time step.
[0050] The reset gate controls the forgetting ratio of the hidden state from the previous time step, determining how much of the hidden state from the previous time step should be forgotten. In other words, it determines how the information from the previous time step is used when calculating the current candidate hidden state. ; in, It is the output of the reset door. It is the weight matrix of the reset gate.
[0051] The candidate hidden state is calculated using the current input and the hidden state of the previous time step. The hidden state of the current time step is obtained by updating the gate output and combining the candidate hidden state with the hidden state of the previous time step. The calculation formula is as follows: ; ; in, It is a candidate hidden state. It is a weight matrix. It is the hidden state at the current time step. It is the hidden state of the previous time step. This is the output of the update gate, and ⊙ represents element-wise multiplication.
[0052] The output of the graph convolutional layer is fed into gated recurrent units (GRUs) step by step. The GRUs process the feature data from the previous time step, pass the resulting hidden state to the next GRU, and then perform calculations with the features from the next time step to obtain a new hidden state. Through iteration, the hidden node of the last time step aggregates the effective information from all previous time steps.
[0053] The Kolmogorov-Arnold network, used as the output layer, possesses powerful nonlinear representation capabilities and feature integration functions. The Kolmogorov-Arnold theorem states that any multivariate continuous function can be represented as a superposition of a finite number of univariate functions: ; in, It is a multivariate continuous function. and Both are unary functions. The Kolmogorov-Arnold network parameterizes these two basis functions through a neural network, and each neuron in the layer uses independent B-spline activation to replace the traditional activation function, forming a learnable activation function.
[0054] B-spline functions are parameterized through a linear combination of B-spline basis functions and trainable coefficients: ; in, For learnable B-spline functions, parameterized as ; For B-spline basis functions, These are trainable coefficients. This represents the number of output nodes in the Kolmogorov-Arnold network layer.
[0055] By stacking, the Kolmogorov-Arnold network can achieve a fully connected network with any number of layers and any number of nodes: ; in, Indicates the first Layer function composition, This is a function compound operator.
[0056] However, the original network requires expanding all intermediate variables to execute different activation functions, leading to severe performance issues. For a layer with an input dimension of m and an output dimension of n, the original implementation needs to expand the input into a tensor of shape (l,n,m) to execute the activation function. However, all activation functions are essentially linear combinations of fixed basis functions. Based on this characteristic, the computation can be restructured as follows: activate the input with different basis functions, and then perform a linear combination.
[0057] Simultaneously, the sparsity handling in the original network needs improvement. The original network's L1 regularization is defined on the input samples, requiring nonlinear operations on tensors of shape (l,n,m), thus incompatible with the reconstruction scheme described above. Therefore, weight-based L1 regularization is used instead, a method more common in neural networks and compatible with the new structure. This reconstruction significantly reduces memory overhead, transforms computation into simple matrix multiplication, and naturally supports both forward and backward propagation.
[0058] The hidden states of the last layer of the gated recurrent unit are used as input to the improved Kolmogorov-Arnold network. Each feature point serves as an initial layer node in the improved Kolmogorov-Arnold network, and is reduced to a one-dimensional representation by several parameterized B-spline functions. These representations are then combined by an external function and passed to the next network layer. Multiple Kolmogorov-Arnold networks are stacked to obtain the final regression prediction.
[0059] S400: Input the standardized dataset and the one-way edge adjacency matrix into the hybrid deep learning prediction model for iterative training. Output the load prediction value and record multiple performance index values in each training round. like Figure 2 As shown, the detailed process of model training and evaluation is as follows: Furthermore, the steps of the cyclic training include: setting an initial number of training rounds, gradually increasing the number of training rounds for cyclic training, until the preset maximum number of training rounds is reached.
[0060] Furthermore, the multiple performance index values include mean absolute error, mean absolute percentage error, root mean square error, and coefficient of determination.
[0061] Specifically, training hyperparameters are set, including learning rate and batch size, and loss function and optimizer are selected. Four metrics are chosen as performance indicators to measure the accuracy of prediction results: mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), and coefficient of determination (R²).
[0062] In this embodiment, preferably, the learning rate is set to 0.001, the sample batch size is set to 32, and the MAE loss function and Adam optimizer are selected.
[0063] MAE, MAPE, and RMSE are evaluation criteria used to reflect the magnitude of the deviation between predicted and actual values; the closer the value is to 0, the smaller the deviation. R² is a statistic used to evaluate the goodness of fit of a regression model, reflecting the degree to which the model explains the variance of the observed data; its maximum value is 1, indicating a perfect fit. The calculation formulas for the above evaluation criteria are shown below: ; ; ; ; in, Indicates the length of the time series. Represents actual data. This is the obtained predicted sequence. This represents the average value of the actual data.
[0064] An initial training epoch value is set, and the number of training epochs is gradually increased for iterative training. At the end of each training session, the epoch number (i.e., the training epoch) is increased by a preset step size until the maximum number of training sessions M is reached. In this embodiment, preferably, the initial training epoch is set to 50, and the step size for each increase is 10. In each training epoch, the model outputs a predicted load value and calculates the four performance metrics mentioned above. The model performance metrics for all training epochs are saved to provide a data foundation for subsequent comprehensive evaluation.
[0065] S500: The SMART method is used to standardize and weight multiple performance index values, select the optimal model parameters based on the comprehensive evaluation index values, and use the optimal model parameters to perform short-term load forecasting of the transformer area based on the data to be predicted.
[0066] Furthermore, the standardization and weight fusion of multiple performance index values using the SMART method includes: The positive and negative performance indicators are standardized separately. Assign temporary weights to each performance metric to improve it from the worst to the best state; Normalize the temporary weights to obtain the weight coefficients of each performance index; The standardized performance index values recorded in each training round are weighted and summed with the corresponding weight coefficients to obtain the comprehensive evaluation index value for each training round. The model parameters corresponding to the training round with the largest comprehensive evaluation index value are selected as the optimal model parameters.
[0067] Specifically, based on the results of cyclic training, there are already existing Each training result sample is composed of [number] training result samples. The performance indicators are composed of, here If the value is 4, then For the first The first sample The values of each indicator are given. Among them, MAE, MAPE, and RMSE are negative indicators, and R² is a positive indicator. The data are standardized to eliminate the influence of different dimensions of the indicators.
[0068] For positive indicators, the standardized formula is: ; For negative indicators, the standardized formula is: ; in, The standardized index value, These are the original index values. Indicates the first training rounds , and They represent the first The maximum and minimum values of each indicator across all samples.
[0069] Determine the criteria for each indicator The relative value score for improving from the worst state (0) to the best state (1). That is, assuming all indicators are at their worst values, i.e. For each indicator Decision-makers set their own Upgraded to Temporary weights This reflects its relative importance.
[0070] Normalize the temporary weights to obtain the weight coefficients for each indicator: ; in, For the first The normalized weighting coefficients of each indicator.
[0071] Using these weights as evaluation index weights, a comprehensive evaluation index for short-term power load forecasting of the distribution area is obtained. : ; in, For the first The comprehensive evaluation index value of each training round, For the first The standardized training round Each performance metric value, For the first The weighting coefficients for each indicator.
[0072] The performance results of all training epochs obtained during the model training phase are evaluated using the aforementioned comprehensive evaluation metrics. For each training epoch, the comprehensive evaluation metric value for that training epoch is obtained by weighting and summing the standardized m performance metric values with their corresponding weight coefficients. .Pick The model parameters corresponding to the training round with the largest value are taken as the optimal model parameters and saved for subsequent use in the task of short-term load forecasting of the distribution area based on the data to be predicted.
[0073] The optimal model parameters selected above are used to perform short-term load forecasting for the transformer substation area. The electrical and environmental characteristic data for the forecast period are preprocessed according to step S200 to obtain a standardized dataset and a one-way edge adjacency matrix. These are then input into a hybrid deep learning prediction model loaded with the optimal model parameters. The model sequentially processes the data through graph convolutional layers, gated recurrent unit layers, and the output layer of an improved Kolmogorov-Arnold network, outputting the load forecast value for that period, thus completing the short-term load forecasting task for the transformer substation area.
[0074] Example 2: This example focuses on a low-voltage distribution transformer area in a residential community of a city. The transformer capacity in this area is 500kVA, serving approximately 280 households. Operational data for 31 consecutive days in October 2024 was collected from this area at a frequency of once every 15 minutes, resulting in 2976 time-step samples. These samples included nine feature dimensions: active power, reactive power, voltage, current, temperature, humidity, wind speed, rainfall, and holiday indicators. Data preprocessing, model building, and training were performed according to the short-term load forecasting method for distribution transformer areas based on integrated diffusion filtering and improved graph convolution of this invention. A comparison of the sequence after noise removal using integrated diffusion filtering technology and the original sequence is shown in the figure below. Figure 3 As shown.
[0075] The SMART method was used to comprehensively evaluate the results of the cyclic training. When the training epoch=150, the comprehensive evaluation index Ω reached its maximum value of 0.907. The corresponding performance indicators of the model were: mean absolute error MAE=17.31kW, mean absolute percentage error MAPE=4.14%, root mean square error RMSE=23.20kW, and coefficient of determination R²=0.9602.
[0076] To verify the technical effectiveness of this invention, comparative experiments were conducted with the traditional LSTM model, the standalone GRU model, and the CNN-LSTM hybrid model under the same dataset and training conditions. The comparison results are shown in Table 1. All comparative experiments were conducted independently, and the result with the best comprehensive evaluation index Ω was selected for presentation. surface Test Comparison Table
[0077] The comparison results show that the MAE of the method in this invention is reduced by 33.3% compared to LSTM, 32.8% compared to GRU, and 19.2% compared to CNN-LSTM; the MAPE is reduced by 34.5%, 33.8%, and 27.6% compared to the above three methods, respectively. The prediction curves of the four methods are compared as follows: Figure 4 As shown.
[0078] The average daily load of this distribution area is approximately 377kW. After adopting the forecasting method of this invention, distribution network dispatchers can accurately predict the peak and valley load conditions 24 hours in advance. In practical applications, based on accurate load forecasting, this distribution area successfully avoided three potential overload risks. Furthermore, by optimizing reactive power compensation configuration, the line loss rate of the distribution area was reduced from the original 4.2% to 3.4%, resulting in monthly electricity savings of approximately 2050kWh.
[0079] This invention effectively removes noise from low-voltage distribution area data through integrated diffusion filtering technology, enhances information interaction between multiple features through graph convolutional layers, and improves the model's ability to fit complex load fluctuations through the learnable B-spline activation function and reconstructed computational structure of the improved Kolmogorov-Arnold network. This significantly improves the accuracy of short-term load forecasting for low-voltage distribution areas and provides effective technical support for the safe and economical operation of the distribution network.
[0080] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. 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 short-term load forecasting of a transformer area based on integrated diffusion filtering and improved graph convolution, characterized in that, The application relates to a short-term load forecasting method based on a hybrid deep learning model. The method comprises the following steps: S100: acquiring electrical characteristic data and environmental characteristic data of a target low-voltage area to form an original characteristic data set; S200: removing noise from the original characteristic data set by using an integrated diffusion filtering technology and standardizing the original characteristic data set to obtain a standardized data set; regarding each time step as a graph node, constructing a single-direction edge adjacency matrix to describe the connection relationship between the nodes, and modeling the power load time series data as graph structure data; S300: constructing a hybrid deep learning prediction model, which comprises a graph convolution layer, a gated recurrent unit layer and an improved Kolmogorov-Arnold network output layer in sequence; S400: inputting the standardized data set and the single-direction edge adjacency matrix into the hybrid deep learning prediction model for cyclic training, outputting a load prediction value and recording a plurality of performance index values in each training round; 2. The method of claim 1, wherein the integrated diffusion filtering and improved graph convolution based feeder short-term load forecasting method is characterized by, S500: standardizing and weight-fusing the plurality of performance index values by using a SMART method, selecting optimal model parameters according to a comprehensive evaluation index value, and using the optimal model parameters to perform short-term load forecasting on to-be-forecasted data of the area. The original characteristic data set comprises: The electrical characteristic data comprises active power, reactive power, voltage value and current value; 3. The integrated diffusion filtering and improved graph convolution based feeder short-term load forecasting method according to claim 1, characterized in that, The environmental characteristic data comprises temperature, humidity, wind speed, rainfall and holiday identification. The step of removing noise from the original characteristic data set by using the integrated diffusion filtering technology comprises: In the forward process, the original load sequence is copied multiple times, and Gaussian noise, uniform noise and Poisson noise are gradually added to the sequence to gradually transition from clean data to high-noise state, so that a sequence representation with different noise levels and different noise types is formed; 4. The method of claim 1, wherein the integrated diffusion filtering and improved graph convolution based feeder short-term load forecasting method is characterized by, In the reverse process, the trained model is used to gradually remove noise from the high-noise representation, and the signal structure is gradually recovered in the reverse direction of the forward process, so that a plurality of smooth load sequences are obtained. The optimal combination of different noise diffusion models is realized by using the integrated idea, and finally the denoised load sequence is obtained.
5. The method of claim 1, wherein the integrated diffusion filtering and improved graph convolution based feeder short-term load forecasting method is characterized by, The standardized processing is performed on the denoised data set by using a maximum-minimum value standardization method.
6. The method of claim 1, wherein the integrated diffusion filtering and improved graph convolution based feeder short-term load forecasting method is characterized by, In the single-direction edge adjacency matrix, information is only allowed to be transmitted from a previous time step node to a subsequent time step node in a single direction. The hybrid deep learning prediction model comprises: The graph convolution layer receives the standardized data set and the single-direction edge adjacency matrix as input, aggregates features through the node connection relationship defined by the adjacency matrix; The gated recurrent unit layer comprises an update gate and a reset gate, the update gate controls the retention proportion of the hidden state of the previous time step, and the reset gate controls the forgetting proportion of the hidden state of the previous time step; 7. The method of claim 1, wherein the integrated diffusion filtering and improved graph convolution based feeder short-term load forecasting method is characterized by, The B-spline function used in the Kolmogorov-Arnold network output layer realizes parameterization through linear combination of B-spline basis functions and trainable coefficients, and reconstructs the calculation process as activating input with different basis functions and then performing linear combination.
8. The method of claim 1, wherein the integrated diffusion filtering and improved graph convolution based feeder short-term load forecasting method is characterized by, The cyclic training step comprises: setting an initial training round, gradually increasing the training round for cyclic training, and stopping until a preset maximum training round is reached. The plurality of performance index values comprise mean absolute error, mean absolute percentage error, root mean square error and determination coefficient.
9. The integrated diffusion filtering and improved graph convolution based feeder short-term load forecasting method of claim 1, wherein, The adopting the SMART method to standardize and weight fuse the performance index values comprises: respectively standardizing the positive performance index and the negative performance index; setting temporary weights for each performance index from the worst state to the optimal state; normalizing the temporary weights to obtain weight coefficients of each performance index; weighting and summing the standardized performance index values recorded in each training round with the corresponding weight coefficients to obtain a comprehensive evaluation index value of each training round, and selecting the corresponding model parameters in the training round with the maximum comprehensive evaluation index value as the optimal model parameters.