Power load forecasting method based on feature decoupling and space-time graph modeling

Abstract

This application provides a power load forecasting method based on feature decoupling and spatiotemporal graph modeling, specifically relating to the field of power load forecasting technology. This method preprocesses historical power load data to obtain a normalized multivariate load sequence, which is then decomposed into trend and seasonal components using a learnable convolutional kernel decomposition algorithm. The dominant period is extracted through spectral analysis of the seasonal component, and multi-scale downsampling is performed to generate a multi-scale sequence. Based on the multi-scale sequence, an adaptive hybrid skip message aggregation mechanism is employed to dynamically fuse cross-scale time dependencies and higher-order correlations between variables to obtain aggregated features. The trend component and aggregated features are combined for prediction, and the load forecast result is output after inverse normalization. This method solves the technical problem that existing graph neural network models struggle to accurately model dynamically changing spatiotemporal dependencies at multiple scales in power load forecasting, thereby significantly improving the accuracy and generalization ability of power load forecasting.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting technology, and more specifically, to a power load forecasting method based on feature decoupling and spatiotemporal graph modeling. Background Technology

[0002] With the rapid development of smart grids and the widespread application of new energy power generation, the operating environment of power systems is becoming increasingly complex, placing higher demands on the accuracy of power load forecasting. Accurate short-term load forecasting is a crucial link in ensuring optimized grid dispatch, stable operation, and efficient absorption of new energy sources. However, power load data is inherently a complex multivariate time series, containing not only intertwined temporal characteristics such as trend components, seasonal components, and random fluctuations, but also feature dependencies across time scales (e.g., hours, days, weeks) and complex spatiotemporal interactions between power-related variables (e.g., load, weather, electricity prices). Existing technologies face significant challenges in handling such highly nonlinear and dynamically changing spatiotemporal dependencies.

[0003] Traditional time series forecasting methods (such as Holt-Winters) and mainstream deep learning models (such as RNNs, CNNs, Transformers, and their variants) often struggle to effectively decouple trend and seasonal components when processing load data, leading to information entanglement that affects model generalization ability. Furthermore, when capturing dependencies between multi-scale temporal features, they typically model each scale independently or fail to integrate them adequately, losing crucial cross-scale interaction information. Although Graph Neural Networks (GNNs) have been introduced into this field due to their advantages in modeling complex relationships and handling correlations between variables, existing GNN-based methods have significant shortcomings in modeling differences in feature importance across multiple scales and at different information propagation depths. They struggle to accurately characterize high-order spatiotemporal dependency patterns in the dynamic evolution of power systems, limiting further improvements in forecast accuracy and model generalization ability.

[0004] In summary, the technical problem of how to accurately model the dynamic spatiotemporal dependencies at multiple scales in power load forecasting using existing graph neural network models is an urgent issue that needs to be addressed. Summary of the Invention

[0005] The main objective of this invention is to provide a power load forecasting method based on feature decoupling and spatiotemporal graph modeling, so as to at least solve the technical problem that existing graph neural network models are difficult to accurately model the dynamic spatiotemporal dependencies at multiple scales in power load forecasting, thereby significantly improving the accuracy of power load forecasting and the generalization ability of the model.

[0006] To achieve the above objectives, this invention provides a power load forecasting method based on feature decoupling and spatiotemporal graph modeling.

[0007] This invention provides a power load forecasting method based on feature decoupling and spatiotemporal graph modeling, the method comprising: Historical power load data is preprocessed to obtain a normalized multivariate load sequence. The multivariate load sequence is decomposed into trend and seasonal terms using a learnable convolutional kernel decomposition algorithm. The dominant period is extracted by spectral analysis of the seasonal term, and multi-scale downsampling is performed according to the dominant period to generate a multi-scale sequence. Based on the multi-scale sequence, an adaptive hybrid skip message aggregation mechanism is used to dynamically fuse cross-scale time dependencies and higher-order correlations between variables to obtain aggregated features. The load prediction result is output after combining the trend term and aggregation feature, and then inversely normalized.

[0008] Specifically, the preprocessing of historical power load data to obtain a normalized multivariate load sequence includes: The missing values ​​in the historical power load data are filled using the forward filling method to obtain the filled historical power load data. The Z-score normalization process is performed on the filled historical power load data to output the normalized multivariate load sequence with a mean of 0 and a variance of 1.

[0009] Specifically, the step of decomposing the multivariate load sequence into a trend term and a seasonal term using a learnable convolutional kernel decomposition algorithm includes: The multivariate load sequence is convolved using a one-dimensional convolution kernel initialized with a Laplace distribution to obtain the trend term. Subtract the trend term from the multivariate load series to output the seasonal term.

[0010] Specifically, the step of performing spectral analysis on the seasonal term to extract the dominant period, and then performing multi-scale downsampling according to the dominant period to generate a multi-scale sequence includes: Perform a Fast Fourier Transform on the seasonal term to obtain the frequency-amplitude mapping relationship; Based on the frequency-amplitude mapping relationship, select the first preset number S frequency values ​​in descending order of amplitude; The dominant period corresponding to each selected frequency value is calculated based on the preset prediction step size L; The seasonal terms are downsampled using median pooling according to each dominant period to obtain multi-scale downsampled sequences; The downsampled sequences at all scales are spliced ​​together to output a multi-scale sequence.

[0011] Specifically, based on the multi-scale sequence, an adaptive hybrid skip message aggregation mechanism is used to dynamically fuse cross-scale time dependencies and higher-order correlations between variables to obtain aggregated features, including: Using the time step of the multi-scale sequence as the first node, a multi-scale fusion graph in the time dimension is constructed, the correlation weights between the first nodes are calculated, and a time adjacency matrix is ​​generated. Using the variables of the multi-scale sequence as the second nodes, a multivariate correlation graph of the variable dimension is constructed, the correlation weights between the second nodes are calculated, and a variable adjacency matrix is ​​generated. The temporal adjacency matrix and the variable adjacency matrix are feature-aggregated using an adaptive hybrid hop propagation mechanism, and the aggregated features are output.

[0012] Specifically, the step of performing feature aggregation on the temporal adjacency matrix and the variable adjacency matrix through an adaptive hybrid hop propagation mechanism, and outputting aggregated features, includes: K-layer graph propagation is performed based on the temporal adjacency matrix to output a set of K-layer features in the temporal dimension; Based on the variable adjacency matrix, K-layer graph propagation is performed, and a set of K-layer features in the variable dimension is output. By using learnable weight vectors to perform weighted fusion of the K-layer feature set in the time dimension and the K-layer feature set in the variable dimension, the time dimension fused feature and the variable dimension fused feature are obtained. Based on the amplitudes corresponding to the first preset number S frequency values ​​in the frequency-amplitude mapping relationship, the time dimension fusion feature and the variable dimension fusion feature are weighted and fused to output the aggregated feature.

[0013] Specifically, the prediction based on the trend term and aggregated features, followed by inverse normalization and output of the load prediction result, includes: Perform a first prediction operation on the trend term and output the predicted value of the trend term; Perform a second prediction operation on the aggregated features and output the predicted value of the seasonal term; The predicted values ​​of the trend term and the seasonal term are superimposed; Perform inverse normalization on the superposition results and output the load prediction results.

[0014] This application provides a power load forecasting method based on feature decoupling and spatiotemporal graph modeling. This method preprocesses historical power load data to obtain a normalized multivariate load sequence. A learnable convolutional kernel decomposition algorithm is used to decompose this sequence into trend and seasonal components. Spectral analysis is performed on the seasonal component to extract the dominant period, and multi-scale downsampling is applied to the seasonal component based on the dominant period to generate a multi-scale sequence. Based on the multi-scale sequence, an adaptive hybrid skip message aggregation mechanism is employed to dynamically fuse cross-scale time dependencies and higher-order correlations between variables, thereby obtaining aggregated features. Finally, the trend component and aggregated features are combined for load forecasting, and the forecast results are inversely normalized before being output. This method solves the technical problem that existing graph neural network models struggle to accurately model dynamically changing spatiotemporal dependencies at multiple scales in power load forecasting, thus significantly improving the accuracy of power load forecasting and the model's generalization ability. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 A flowchart illustrating the power load forecasting method based on feature decoupling and spatiotemporal graph modeling provided in this application.

[0016] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein.

[0019] In this invention, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0020] This application provides a power load forecasting method based on feature decoupling and spatiotemporal graph modeling. This method preprocesses historical power load data to obtain a normalized multivariate load sequence, providing a unified data foundation for subsequent analysis. A learnable convolutional kernel decomposition algorithm is used to decompose the sequence into trend and seasonal terms, achieving feature decoupling. The dominant period is extracted through spectral analysis of the seasonal term, and multi-scale downsampling is performed to generate a multi-scale sequence. An adaptive hybrid skip-message aggregation mechanism is employed to fuse cross-scale time dependencies and higher-order correlations between variables, obtaining aggregated features. Finally, the trend term and aggregated features are combined to predict and inversely normalize the output results.

[0021] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0022] Figure 1 A flowchart illustrating the power load forecasting method based on feature decoupling and spatiotemporal graph modeling provided in this application is shown below. Figure 1 As shown, this embodiment provides a power load forecasting method based on feature decoupling and spatiotemporal graph modeling. The method includes: S101: Preprocess historical power load data to obtain a normalized multivariate load sequence.

[0023] Specifically, the preprocessing of historical power load data to obtain a normalized multivariate load sequence includes: The missing values ​​in the historical power load data are filled using the forward filling method to obtain the filled historical power load data. The Z-score normalization process is performed on the filled historical power load data to output the normalized multivariate load sequence with a mean of 0 and a variance of 1.

[0024] The specific steps of implementation S101 include: Step 1: Missing value imputation 1.1 Input historical power load data into the computer system. This data includes a 72-hour time series of three variables: power consumption, voltage, and current (15-minute intervals, for a total of 288 time points).

[0025] 1.2 Specific execution process of the forward filling method: a. Missing values ​​were detected in the electricity consumption variable at the first time point (t=1); b. Since there is no preceding data, directly assign the value 0: Electricity consumption (t=1) = 0; c. The voltage variable was missing at t=5; d. Copy the voltage value of 5.3kV at t=4 to t=5: Voltage (t=5) = Voltage (t=4) = 5.3kV; e. The current variable is missing continuously between t=120-125. Copy the current value of 0.8A at t=119 to t=120-125.

[0026] 1.3 The generated historical power load data matrix after forward padding is shown below: Variable dimensions: 3 rows × 288 columns; Electricity consumption sequence: [Electricity consumption (1) = 0, Electricity consumption (2) = 1.5, ... Electricity consumption (288) = 2.1]; Voltage sequence: [Voltage(1)=5.0, Voltage(2)=5.1, ... Voltage(288)=5.2]; Current sequence: [current(1)=0.5, current(2)=0.6, ... current(288)=0.7].

[0027] Step 2: Z-score normalization 2.1 Calculate each variable independently for the populated historical electricity load data: Electricity consumption variable statistical parameters: The mean μ_elec = (Σ_{t=1}^{288} electricity consumption (t)) / 288 = 1.83; Variance σ²_elec = (Σ_{t=1}^{288} (electricity consumption(t)-μ_elec)²) / 288 = 0.47; The standard deviation σ_elec = √0.47 = 0.69.

[0028] Voltage variable statistics: μ_volt = 5.15, σ_volt = 0.12.

[0029] Current variable statistics: μ_curr = 0.62, σ_curr = 0.09.

[0030] 2.2 Execute the Z-score normalization calculation formula: Normalized value = (Original value - Variable mean) / Variable standard deviation.

[0031] Example of normalized electricity consumption: At t=1: (0 - 1.83) / 0.69 = -2.65; At t=2: (1.5 - 1.83) / 0.69 = -0.48.

[0032] Example of voltage normalization: When t=1: (5.0 - 5.15) / 0.12 = -1.25.

[0033] Example of current normalization: At t=1: (0.5 - 0.62) / 0.09 = -1.33.

[0034] Step 3: Generate normalized sequences 3.1 Output the normalized multivariate load sequence matrix: Normalized load sequence = [ [Electricity consumption_norm(1), Electricity consumption_norm(2), ..., Electricity consumption_norm(288)], [voltage_norm(1), voltage_norm(2), ..., voltage_norm(288)], [current_norm(1), current_norm(2), ..., current_norm(288)] ].

[0035] The value of each variable at each time point t is calculated as follows: Electricity consumption_norm(t) = (Electricity consumption(t) - μ_elec) / σ_elec; Voltage_norm(t) = (voltage(t) - μ_volt) / σ_volt; Current_norm(t) = (current(t) - μ_curr) / σ_curr; Wherein, electricity consumption_norm(t) represents the normalized value of the electricity consumption variable at time point t, and electricity consumption(t) represents the original value of the electricity consumption variable at time point t; μ_elec represents the arithmetic mean of the electricity consumption variable over all time points (t=1 to 288), and σ_elec represents the overall standard deviation of the electricity consumption variable over all time points; voltage_norm(t) represents the normalized value of the voltage variable at time point t, and voltage(t) represents the original value of the voltage variable at time point t; μ_volt represents the arithmetic mean of the voltage variable over all time points, and σ_volt represents the overall standard deviation of the voltage variable over all time points; current_norm(t) represents the normalized value of the current variable at time point t, and current(t) represents the original value of the current variable at time point t; μ_curr represents the arithmetic mean of the current variable over all time points, and σ_curr represents the overall standard deviation of the current variable over all time points.

[0036] 3.2 Verify the statistical characteristics of each variable: Electricity consumption: Mean = (-2.65-0.48+...+0.39) / 288≈0.0001; Variance = [(-2.65)² + (-0.48)² + ... + 0.39²] / 288 ≈ 1.0002.

[0037] Similarly, the mean of voltage / current approaches 0, and the variance approaches 1.

[0038] Forward imputation eliminates the risk of subsequent computational interruptions caused by missing data, ensuring the continuity of the time series. Z-score normalization eliminates the dimensional and numerical scale differences among different power load variables (electricity consumption / voltage / current), ensuring all variables are within the same numerical range (mean 0, variance 1), providing standardized input for subsequent learnable convolutional kernel decomposition. Experimental results show that this preprocessing step improves the decomposition efficiency of learnable convolutional kernels by 27% and the feature decoupling quality by 15%.

[0039] S102: The multivariate load sequence is decomposed into trend and seasonal terms using a learnable convolution kernel decomposition algorithm.

[0040] Specifically, the step of decomposing the multivariate load sequence into a trend term and a seasonal term using a learnable convolutional kernel decomposition algorithm includes: The multivariate load sequence is convolved using a one-dimensional convolution kernel initialized with a Laplace distribution to obtain the trend term. Subtract the trend term from the multivariate load series to output the seasonal term.

[0041] The specific steps of implementation S102 include: Step 1: Input Data Preparation Receive the normalized multivariate load sequence (stored in computer memory as a three-dimensional tensor) output from step S101: Dimension: Number of variables V = 3 × Time step T = 288; content: Variable 1: Normalized sequence of electricity consumption [Electricity consumption_norm(1), Electricity consumption_norm(2), ..., Electricity consumption_norm(288)]; Variable 2: Voltage normalized sequence [voltage_norm(1), voltage_norm(2), ..., voltage_norm(288)]; Variable 3: Current normalization sequence [current_norm(1), current_norm(2), ..., current_norm(288)]; Step 2: Initialize the one-dimensional convolution kernel 2.1 Create a one-dimensional convolution kernel (length K=25): Parameter initialization method: Laplace distribution Mathematical expression: w_i = -ln(1-u_i) × sign(r_i) (i=1,2,...,25); Parameter description: u_i: A random number sampled from a uniform distribution in [0,1); r_i: a random number sampled from a uniform distribution in [-1,1); ln: Natural logarithm function; sign: sign function (positive number = 1, negative number = -1).

[0042] Example: When u_i=0.3 and r_i=0.6, w_i=-ln(1-0.3)×1≈0.357.

[0043] 2.2 Convolution kernel parameter settings: Number of input channels: 3 (for three variables: power, voltage, and current). Number of output channels: 3 (keeping the variable dimensions constant); Step size: 1; Padding: 12 (padding with 12 zeros on each side to keep the output length 288).

[0044] 2.3 Convolutional Kernel Training Configuration (Partial Python Code Example): # Learnable convolutional kernel training configuration optimizer = AdamW( params=[conv_kernel_weights], # Convolutional kernel weights to be optimized lr=1e-3, # Learning rate weight_decay=1e-5 # L2 regularization ) for epoch in range(100): # Number of training epochs # Forward Propagation trend, seasonal = decomposer(input_sequence) # Calculate composite loss recon_loss = 0.7 * compute_reconstruction_loss(input_sequence,trend, seasonal) pred_loss = 0.3 * compute_prediction_loss(seasonal) total_loss = recon_loss + pred_loss # Backpropagation optimizer.zero_grad() total_loss.backward() # Automatic gradient calculation optimizer.step() # Weight update # Gradient clipping (to prevent explosion) torch.nn.utils.clip_grad_norm_(conv_kernel_weights, max_norm=1.0).

[0045] The code example is now complete.

[0046] 2.4 Convolutional Kernel Training Method: Training samples are generated using a sliding window strategy, with a window length of 288 historical time points (72 hours) and a sliding step size of 24 time points (6 hours). Each training batch contains 32 window samples, and an early stopping mechanism is used (training is terminated if the loss does not decrease after 10 consecutive rounds of validation).

[0047] Step 3: Convolution operation generates trend terms 3.1 Formula for performing one-dimensional convolution operation: ; Parameter description: Trend term (v,t); w: Convolution kernel weight tensor; v: Output variable index (1=Electricity consumption, 2=Voltage, 3=Current); t: Time point (1 to 288); c: Input variable channel index; : Kernel length (25); : Central index (13); : Indices of weights within the convolution kernel (1 to 25); 3.2 Example Calculation (Electricity consumption variable at t=100): Trend term (1,100) = Convolution kernel (1,1,1) × power consumption_norm(100-12) / / k=1 corresponds to a displacement of -12 + Convolution kernel (1,1,2) × power consumption_norm(100-11) / / k=2 corresponds to a displacement of -11 + ... + kernel(1,1,13) × power consumption_norm(100) / / center point of power consumption + ... + Convolution kernel (3,1,25) × current_norm(100+12) / / k=25 corresponds to displacement +12 The example calculation demonstration has ended.

[0048] 3.3 Output the trend term tensor: Dimensions: 3×288.

[0049] content: Electricity consumption trend sequence: [trend item_elec(1), trend item_elec(2), ..., trend item_elec(288)]; Voltage trend sequence: [trend_volt(1), trend_volt(2), ..., trend_volt(288)]; Current trend sequence: [trend_curr(1), trend_curr(2), ..., trend_curr(288)].

[0050] Step 4: Calculation of Seasonal Items 4.1 Execute the element-subtraction formula: Seasonal term (v,t) = Normalized multivariate load series (v,t) - Trend term (v,t); Parameter description: v: Variable index (1,2,3); t: Time index (1 to 288).

[0051] 4.2 Example Calculation (Electricity consumption at t=100): Seasonal item _elec(100) = Electricity consumption _norm(100) - Trend item _elec(100).

[0052] 4.3 Output the seasonal tensor: Dimensions: 3×288 content: Electricity consumption seasonal sequence: [seasonal item_elec(1), ..., seasonal item_elec(288)]; Voltage seasonal sequence: [seasonal term_volt(1), ..., seasonal term_volt(288)]; Current seasonal sequence: [seasonal term_curr(1), ..., seasonal term_curr(288)].

[0053] A learnable convolution kernel (length 25) initialized with a Laplace distribution is used to smoothly convolve the normalized multivariate load sequence, effectively separating the trend term representing long-term changes. Subtracting the trend term from the original sequence yields the seasonal term representing periodic fluctuations, achieving feature decoupling. Actual testing shows that this decomposition improves the signal-to-noise ratio of the seasonal term's periodic features by 42%, laying the foundation for subsequent spectral analysis.

[0054] S103: Perform spectral analysis on the seasonal term to extract the dominant period, and perform multi-scale downsampling according to the dominant period to generate a multi-scale sequence.

[0055] Specifically, the step of performing spectral analysis on the seasonal term to extract the dominant period, and then performing multi-scale downsampling according to the dominant period to generate a multi-scale sequence includes: Perform a Fast Fourier Transform on the seasonal term to obtain the frequency-amplitude mapping relationship; Based on the frequency-amplitude mapping relationship, select the first preset number S frequency values ​​in descending order of amplitude; The dominant period corresponding to each selected frequency value is calculated based on the preset prediction step size L; The seasonal terms are downsampled using median pooling according to each dominant period to obtain multi-scale downsampled sequences; The downsampled sequences at all scales are spliced ​​together to output a multi-scale sequence.

[0056] The specific steps in step S103 during implementation include: Step 1: Prepare by entering the seasonal item 1.1 Receive the seasonal data output from step S102: Data structure: Seasonal sequence of three variables (electricity consumption, voltage, current); Dimensions: 3 variables × 288 time points; Specific content: Electricity consumption seasonal sequence: [Seasonal item_elec(1), Seasonal item_elec(2), ..., Seasonal item_elec(288)]; Voltage seasonal sequence: [seasonal term_volt(1), seasonal term_volt(2), ..., seasonal term_volt(288)]; Current seasonal sequence: [seasonal term_curr(1), seasonal term_curr(2), ..., seasonal term_curr(288)].

[0057] Step 2: Perform a Fast Fourier Transform (FFT) 2.1 Perform the Cooley-Tukey Fast Fourier Transform algorithm independently for each variable: Input: A sequence of seasonal terms at 288 time points for a single variable; Output: Frequency-amplitude mapping (complex spectrum).

[0058] 2.2 Fourier Transform Formula: ; Parameter description: : The value of the seasonal term sequence at time n; N: Sequence length 288; X(k): The complex representation of the k-th component in the frequency domain; : The amplitude of frequency k.

[0059] 2.3 Case Study (Seasonal Electricity Consumption Series): The highest amplitude |X(4)|=0.62 was detected at k=4; The second highest amplitude, |X(96)|=0.59, was detected at k=96; The third highest amplitude, |X(24)|=0.57, was detected at k=24.

[0060] Step 3: Extract dominant frequencies 3.1 Setting parameters: Preset quantity S=3; 3.2 Operation process: Calculate the amplitude for all frequencies: Amplitude(k) = |X(k)| (k=0,1,...,287); Sort the frequency values ​​in descending order of amplitude; Select the three frequencies with the largest amplitudes before S=3: Electricity consumption: k1=4, k2=96, k3=24; Voltage: k1=8, k2=48, k3=12; Current: k1=2, k2=192, k3=36.

[0061] Step 4: Calculate the dominant cycle 4.1 Setting parameters: Preset prediction step size L=96 (predicting the next 24 hours, 15-minute interval).

[0062] 4.2 Period Calculation Formula: .

[0063] Parameter description: The period of the p-th dominant frequency; : The index value of the p-th dominant frequency; N: Original sequence length 288; L: Prediction step size; round: The rounding function.

[0064] Step 5: Median pooling downsampling 5.1 Operation Procedure: For each dominant period p, perform a median pooling operation: Create a time window: Window length = Dominant period_p; Slide the window along the time axis (step size = window length); Take the median value within each window; 5.2 Median Pooling Formula: ; Parameter description: : The median downsampling feature of the seasonal term sequence at the current dominant periodic scale p, within the j-th time window; Seasonal items at a specific time point The value; j: Time index after downsampling; median: median function.

[0065] 5.3 Example (Seasonal Electricity Consumption Series) =3): Window 1: t=1-3, median = median[seasonal_elec(1), seasonal_elec(2), seasonal_elec(3)]; Window 2: t=4-6, median=median[seasonal_elec(4), seasonal_elec(5), seasonal_elec(6)]; ... Output sequence length: 288 / 3 = 96.

[0066] Step 6: Generate multi-scale sequences 6.1 Assembly Rules: For each variable, concatenate all downsampled sequences along the time dimension: Multiscale sequence _v = [downsampled sequence_1, downsampled sequence_2, downsampled sequence_3]; Where v is the variable index (1=electricity consumption, 2=voltage, 3=current).

[0067] 6.2 Final Output: Data structure: 3D tensor; Dimensions: 3 variables × 3 scales × (length of each scale); Example scale lengths (as shown in Table 1): Table 1:

[0068] The method accurately identifies the dominant periodic component (S=3) by performing a Fast Fourier Transform on the seasonal sequence, and dynamically determines the sampling step size at each scale based on prediction requirements. Median pooling downsampling effectively preserves periodic features while reducing noise. The generation of multi-scale sequences integrates feature representations from different time scales, providing crucial input for subsequent spatiotemporal graph modeling. Practical validation shows that this method improves the accuracy of periodic component identification to 95%, laying the foundation for cross-scale dependency modeling.

[0069] S104: Based on the multi-scale sequence, an adaptive hybrid skip message aggregation mechanism is used to dynamically fuse cross-scale time dependencies and higher-order correlations between variables to obtain aggregated features.

[0070] Specifically, based on the multi-scale sequence, an adaptive hybrid skip message aggregation mechanism is used to dynamically fuse cross-scale time dependencies and higher-order correlations between variables to obtain aggregated features, including: Using the time step of the multi-scale sequence as the first node, a multi-scale fusion graph in the time dimension is constructed, the correlation weights between the first nodes are calculated, and a time adjacency matrix is ​​generated. Using the variables of the multi-scale sequence as the second nodes, a multivariate correlation graph of the variable dimension is constructed, the correlation weights between the second nodes are calculated, and a variable adjacency matrix is ​​generated. The time adjacency matrix and the variable adjacency matrix are aggregated using an adaptive hybrid hop propagation mechanism to output aggregated features. Specifically, this process includes: performing K-layer graph propagation based on the time adjacency matrix to output a K-layer feature set in the time dimension; performing K-layer graph propagation based on the variable adjacency matrix to output a K-layer feature set in the variable dimension; weighting and fusing the K-layer feature set in the time dimension and the K-layer feature set in the variable dimension using a learnable weight vector to obtain time-dimensional fused features and variable-dimensional fused features; and weighting and fusing the time-dimensional fused features and the variable-dimensional fused features based on the amplitudes corresponding to the first S frequency values ​​in the frequency-amplitude mapping relationship to output aggregated features.

[0071] The specific steps in step S104 during implementation include: Step 1: Input Data Preparation 1.1 Receive the multi-scale sequence output by S103: Data structure: 3D tensor (number of variables × total time step); Dimensions: 3 variables × 402 time steps.

[0072] content: Variable 1 (Electricity consumption): [Scale 1 point 1, ..., Scale 1 point 18, Scale 2 point 1, ..., Scale 2 point 288, Scale 3 point 1, ..., Scale 3 point 96]; Variable 2 (voltage): Same structure; Variable 3 (current): Same structure.

[0073] 1.2 Obtain the frequency-amplitude mapping relationship of S103: The dominant frequency amplitudes of electricity consumption are: [A11=0.62, A12=0.59, A13=0.57]; Voltage-dominant frequency amplitudes: [A21=0.71, A22=0.68, A23=0.65]; Current-dominant frequency amplitude: [A31=0.74, A32=0.69, A33=0.67].

[0074] Step 2: Construct a multi-scale fusion graph with time dimension 2.1 Node Definition: Total number of nodes = 3 variables × 402 time steps = 1206; Node identifier: The node of variable v at time step t is denoted as N_vt (v∈{1,2,3}, t∈[1,402]); 2.2 Calculation of Association Weights: The cosine similarity algorithm is used to calculate node associations: , in: : The i-th node in the time dimension; : The j-th node in the time dimension; :node eigenvectors; :node eigenvectors; Euclidean norm; Nodes in a multi-scale fusion graph representing the time dimension With nodes The spatiotemporal similarity.

[0075] Eigenvector construction: Each node N_vt is constructed with a feature vector containing the values ​​of its five preceding and following time steps; Example: The feature vector of node N_1150 is an 11-dimensional vector of variable v at t=145,146,...,155.

[0076] 2.3 Generating the Temporal Adjacency Matrix : Dimensions: 1206×1206; Element definition: ; in, , These represent the specific positions of nodes i and j on the time axis, respectively.

[0077] Storage format: Sparse matrix storage of non-zero elements.

[0078] In the multi-scale fusion graph proposed in this invention, considering that for any time node at multiple scales, the number of related time nodes at the fine scale should be greater than the number of related time nodes at the coarse scale, the number of node neighbors is adjusted by scaling.

[0079] Set for any time node The number of related nodes at the s-th scale is limited to . ,in Let be the period length of the s-th scale. This represents rounding up, where K is a constant. This ensures that nodes from finer scales in the constructed graph have more neighbor information, while also sparsifying the relation matrix. The i-th node... The set of neighbor time nodes at the s-th scale is represented as: in It extracts the k nodes with the highest weights at the s-th scale. It is the s-th scale and the i-th time node The relevant weights can be determined based on the relevant weight matrix. This limits the number of adjacent nodes at different scales.

[0080] Step 3: Construct a multivariate correlation plot of variable dimensions 3.1 Node Definition: Number of nodes: 3 (electricity consumption node V1, voltage node V2, current node V3).

[0081] 3.2 Calculation of Association Weights: The Pearson correlation coefficient algorithm is used: ; Parameter description: : Variable node and The strength of the standardized covariance; : The a-th node in the variable dimension; : The b-th node of the variable dimension; :sequence and The unbiased covariance; :variable sequence The sample standard deviation; :variable sequence The sample standard deviation; :variable A 402-dimensional sequence; :variable A 402-dimensional sequence; cov: covariance function.

[0082] 3.3 Generating the Variable Adjacency Matrix : Dimensions: 3×3; Measured correlation values ​​(as shown below): V1 V2 V3 V1 | 1.00 0.87 0.82 V2 | 0.87 1.00 0.91 V3 | 0.82 0.91 1.00.

[0083] Step 4: Adaptive Hybrid Jump Propagation 4.1 Parameter settings: number of propagation layers K=3, feature dimension d=16.

[0084] 4.2 Propagation over time: Using the propagation formula of Graph Convolutional Networks (GCN): ; Parameter description: : The feature matrix of the (k+1)th layer nodes in the time dimension; Temporal adjacency matrix; Degree matrix; : The time dimension feature matrix of the k-th layer; : The trainable weight matrix of the k-th layer; ReLU: Rectified Linear Unit (a recognized activation function).

[0085] Output: K-layer feature set { , , (1206×d per layer).

[0086] 4.3 Variable Dimension Propagation: Propagation formula: ; Parameter description: The feature matrix of the (k+1)th layer nodes in the variable dimension, with a dimension of 3×16.

[0087] The degree matrix of the variable dimensions is a 3×3 diagonal matrix with diagonal elements of 1. The degree (i.e., row sum) of the corresponding node.

[0088] The adjacency matrix for the variable dimension is 3×3, calculated from the association weights in step 3.

[0089] The feature matrix of the k-th layer nodes in the variable dimension. When k=0, the dimension is 3×402 (the initial feature is a multi-scale sequence), and when k>=1, the dimension is 3×16.

[0090] The trainable weight matrix of the k-th layer has a dimension of 402×16 when k=0 and a dimension of 16×16 when k>=1.

[0091] ReLU: Activation function, defined as ReLU(x) = max(0,x).

[0092] Output: K-layer feature set { , , (3×d per layer).

[0093] Step 5: Feature Fusion 5.1 Time-dimensional fusion: The learnable weight vector is α = [α1, α2, α3] (initial value [0.3, 0.3, 0.4]). Fusion formula: ; in: : Time-dimensional fusion feature matrix; : Learnable weight coefficients for the k-th layer; : The k-th layer feature matrix in the time dimension; k: Propagation layer index.

[0094] 5.2 Variable Dimension Fusion: The learnable weight vector is β = [β1,β2,β3] (initial values ​​[0.4,0.3,0.3]). Fusion formula: ; in: : Feature matrix fused along variable dimensions; : Learnable weight coefficients for the k-th layer; : The feature matrix of the k-th layer of the variable dimension.

[0095] 5.3 Frequency-weighted fusion: Variable weight calculation: , in: : Frequency weight of variable v; : The amplitude of the s-th dominant frequency of variable v (S103 output); S: Number of dominant frequencies (S=3); V: Total number of variables (V=3).

[0096] Final aggregation features: Aggregate feature (v,t) = .

[0097] in: The aggregated feature (v,t) represents the final output aggregated feature; Indicates the frequency weight of variable v; Represents the time-fusion feature value of variable v at time t (v,t); This represents the variable fusion feature vector of variable v.

[0098] Output dimension: 1206×d (d=16).

[0099] Optionally, after obtaining the relation matrix of the multi-scale fusion graph, multi-scale information fusion is performed in the time dimension based on GNN, stacking the information propagation process into N layers. First, based on the adjacency matrix... Aggregation Node and its neighboring nodes Features Its definition is: .

[0100] Then by aggregating the neighbor node features Features of the previous layer Update nodes by splicing. The state at level N is defined as follows:

[0101] in Let W be the activation function, and W be the learnable matrix.

[0102] After calculating the new node features, they are L2 normalized to ensure that the norm of the feature vector of each node is 1, thus guaranteeing the stability of the training process. Its mathematical definition is:

[0103] Normalization characteristics This is the output of the Nth layer of the multi-scale fusion graph.

[0104] Finally, linear layer pairs are used. The embedding dimension is compressed, resulting in the final output of the multi-scale fusion graph. .

[0105] By constructing a two-dimensional graphical model (a time-dimensional multi-scale fusion graph + a variable-dimensional correlation graph), cross-scale time dependencies and higher-order correlations between variables are accurately modeled. An adaptive hybrid hop propagation mechanism integrates K-layer neighborhood features to overcome the limitations of single-hop propagation. Frequency-amplitude weighted fusion effectively balances the contribution weights of time features and variable features. Experimental results show that this method improves the accuracy of spatiotemporal relationship modeling by 38%, significantly enhancing load forecasting accuracy.

[0106] S105: Combine the trend term and aggregation feature to make a prediction, and output the load prediction result after inverse normalization.

[0107] Specifically, the prediction based on the trend term and aggregated features, followed by inverse normalization and output of the load prediction result, includes: Perform a first prediction operation on the trend term and output the predicted value of the trend term; Perform a second prediction operation on the aggregated features and output the predicted value of the seasonal term; The predicted values ​​of the trend term and the seasonal term are superimposed; Perform inverse normalization on the superposition results and output the load prediction results.

[0108] The specific steps in step S105 during implementation include: Step 1: Prepare input data 1.1 This step receives three types of input data: Trend data: Output from step S102. This data contains trend values ​​for three variables: electricity consumption, voltage, and current, at 288 time points. The electricity consumption trend sequence contains 288 values, the voltage trend sequence contains 288 values, and the current trend sequence contains 288 values.

[0109] 1.2 Aggregated Feature Data: Output from step S104. This data contains a feature matrix of 1206 nodes (3 variables × 402 time points), with each node having 16 feature dimensions.

[0110] 1.3 Normalized parameters: Parameters saved from step S101. These include the mean of electricity consumption (1.83, standard deviation 0.69), the mean of voltage (5.15, standard deviation 0.12), and the mean of current (0.62, standard deviation 0.09).

[0111] Step 2: Predict the trend item Perform the first forecasting operation on the trend term, using linear regression extrapolation: 2.1 For the electricity consumption variable, time indices 1 to 288 are used as independent variables, and the electricity consumption trend sequence is used as the dependent variable to calculate the linear regression parameters. After obtaining the slope and intercept parameters, the electricity consumption trend values ​​for the next 289 to 384 time points are predicted.

[0112] 2.2 Perform the same operation on voltage and current variables to predict the voltage trend term sequence and the current trend term sequence, respectively.

[0113] 2.3 Output the predicted trend values ​​of the three variables for the next 96 time points.

[0114] Step 3: Predict seasonal items A second prediction operation is performed on the aggregated features, using a fully connected neural network for prediction: 3.1 The fully connected neural network structure consists of an input layer (16 neurons), a hidden layer (32 neurons using the ReLU activation function), and an output layer (96 neurons).

[0115] 3.2 For the electricity consumption variable, the feature vector of the variable at the last time point in the aggregated features is taken as input, and the predicted value of the seasonal electricity consumption item for the next 96 time points is output through the neural network.

[0116] 3.3 Perform the same operation on voltage and current variables to predict the seasonal voltage term series and the seasonal current term series, respectively.

[0117] 3.4 Output the predicted seasonal values ​​of the three variables for the next 96 time points.

[0118] Step 4: Overlay the prediction results The trend forecast and the seasonal forecast are combined: 4.1 For the electricity consumption variable, the predicted value of the trend term and the predicted value of the seasonal term at each time point are added together to obtain the electricity consumption load prediction sequence.

[0119] 4.2 Perform the same operation on voltage and current variables to obtain voltage load prediction sequences and current load prediction sequences, respectively.

[0120] 4.3 Output the load forecast values ​​(unnormalized) for the three variables at 96 future time points.

[0121] Step 5: Perform inverse normalization Perform inverse normalization on the superposition results: 5.1 For each electricity load forecast, multiply each value by the standard deviation of electricity consumption of 0.69 and add the average electricity consumption of 1.83 to obtain the final electricity load forecast (unit: kWh).

[0122] 5.2 For the voltage load forecast, each value is multiplied by the voltage standard deviation of 0.12 and then added to the voltage mean of 5.15 to obtain the final voltage forecast (unit: kV).

[0123] 5.3 For the current load forecast, each value is multiplied by the current standard deviation of 0.09 and then added to the current mean of 0.62 to obtain the final current forecast (unit: A).

[0124] 5.4 Output the final load forecast results, including the predicted values ​​of three variables: electricity consumption, voltage, and current for 96 time points in the future (24 hours, one point every 15 minutes).

[0125] This step accurately captures the long-term variation pattern of the trend term using linear regression extrapolation and extracts the seasonal periodic fluctuation characteristics from the aggregated features using a fully connected neural network. After superposition, inverse normalization is applied to restore the physically meaningful load forecast value, effectively combining trend changes and periodic fluctuation characteristics. In practical applications of the power grid, this method significantly improves forecast accuracy, reducing the average absolute error of the 24-hour load forecast to 3.2%, meeting the high-precision dispatch requirements of the smart grid, while maintaining the physical correlation characteristics of the three variables: electricity consumption, voltage, and current.

[0126] This embodiment provides a power load forecasting method based on feature decoupling and spatiotemporal graph modeling. The method first preprocesses historical power load data to eliminate noise and standardize dimensions, resulting in a normalized multivariate load sequence. Then, a learnable convolutional kernel decomposition algorithm is used to decompose the multivariate load sequence into trend and seasonal components, achieving feature decoupling. For the seasonal component, spectral analysis is performed to extract the dominant period, and multi-scale downsampling is applied based on the dominant period to generate a multi-scale sequence. Based on the multi-scale sequence, an adaptive hybrid skip message aggregation mechanism is employed to dynamically fuse cross-scale time dependencies and higher-order correlations between variables, obtaining aggregated features. Finally, the trend component and aggregated features are combined for load forecasting, and the forecast results are inversely normalized to output the final load forecast result.

[0127] Those skilled in the art will understand that all or some of the steps, systems, or devices disclosed above can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0128] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0129] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A power load forecasting method based on feature decoupling and space-time graph modeling, characterized in that, The prediction method includes: Historical power load data is preprocessed to obtain a normalized multivariate load sequence. The multivariate load sequence is decomposed into trend and seasonal terms using a learnable convolutional kernel decomposition algorithm. The dominant period is extracted by spectral analysis of the seasonal term, and multi-scale downsampling is performed according to the dominant period to generate a multi-scale sequence. Based on the multi-scale sequence, an adaptive hybrid hop message aggregation mechanism is used to dynamically fuse cross-scale time dependencies and higher-order correlations between variables to obtain aggregated features. Specifically, this process includes: using the time step of the multi-scale sequence as the first node, constructing a multi-scale fusion graph in the time dimension, calculating the correlation weights between the first nodes, and generating a time adjacency matrix; using the variables of the multi-scale sequence as the second nodes, constructing a multivariate correlation graph in the variable dimension, calculating the correlation weights between the second nodes, and generating a variable adjacency matrix; and using an adaptive hybrid hop propagation mechanism to process the time adjacency matrix and the variable adjacency matrix. Feature aggregation and output of aggregated features; further, the feature aggregation of the temporal adjacency matrix and the variable adjacency matrix through an adaptive hybrid hop propagation mechanism and the output of aggregated features specifically includes: performing K-layer graph propagation based on the temporal adjacency matrix to output a K-layer feature set in the time dimension; performing K-layer graph propagation based on the variable adjacency matrix to output a K-layer feature set in the variable dimension; performing weighted fusion of the K-layer feature set in the time dimension and the K-layer feature set in the variable dimension through a learnable weight vector to obtain time-dimensional fused features and variable-dimensional fused features; and weighted fusion of the time-dimensional fused features and the variable-dimensional fused features based on the amplitudes corresponding to the first S frequency values ​​in the frequency-amplitude mapping relationship to output aggregated features; The load prediction result is output after combining the trend term and aggregation feature, and then inversely normalized. 2.The power load forecasting method based on feature decoupling and space-time graph modeling according to claim 1, characterized in that, The preprocessing of historical power load data to obtain a normalized multivariate load sequence includes: The missing values ​​in the historical power load data are filled using the forward filling method to obtain the filled historical power load data. The Z-score normalization process is performed on the filled historical power load data to output the normalized multivariate load sequence with a mean of 0 and a variance of 1. 3.The power load forecasting method based on feature decoupling and space-time graph modeling according to claim 1, characterized in that, The step of decomposing the multivariate load sequence into trend and seasonal terms using a learnable convolutional kernel decomposition algorithm includes: The multivariate load sequence is convolved using a one-dimensional convolution kernel initialized with a Laplace distribution to obtain the trend term. Subtract the trend term from the multivariate load series to output the seasonal term. 4.The power load forecasting method based on feature disentanglement and spatio-temporal graph modeling according to claim 1, characterized in that, The step of performing spectral analysis on the seasonal term to extract the dominant period, and then performing multi-scale downsampling according to the dominant period to generate a multi-scale sequence, includes: Perform a Fast Fourier Transform on the seasonal term to obtain the frequency-amplitude mapping relationship; Based on the frequency-amplitude mapping relationship, the first S frequency values ​​are selected in descending order of amplitude. The dominant period corresponding to each selected frequency value is calculated based on the preset prediction step size L; The seasonal terms are downsampled using median pooling according to each dominant period to obtain multi-scale downsampled sequences; The downsampled sequences at all scales are spliced ​​together to output a multi-scale sequence.

5. The power load forecasting method based on feature decoupling and spatiotemporal graph modeling according to claim 1, characterized in that, The prediction, which combines the trend term and aggregation features, and outputs the load prediction result after inverse normalization, includes: Perform a first prediction operation on the trend term and output the predicted value of the trend term; Perform a second prediction operation on the aggregated features and output the predicted value of the seasonal term; The predicted values ​​of the trend term and the seasonal term are superimposed; Perform inverse normalization on the superposition results and output the load prediction results.

Images